Stop Building Bots, Start Building Runtimes: A Field Guide to Microsoft Agents

1
00:00:00,000 --> 00:00:03,000
Everyone is calling Build 2026, the AI event.

2
00:00:03,000 --> 00:00:05,400
They're focused on the keynotes and the new models.

3
00:00:05,400 --> 00:00:07,000
But the real shift is much quieter.

4
00:00:07,000 --> 00:00:08,760
It isn't happening in the headlines.

5
00:00:08,760 --> 00:00:10,400
It's happening in the structure underneath.

6
00:00:10,400 --> 00:00:11,680
But here's what actually changed.

7
00:00:11,680 --> 00:00:14,520
We moved from chatbots that assist to agents that operate.

8
00:00:14,520 --> 00:00:16,440
That sounds like a small feature update.

9
00:00:16,440 --> 00:00:17,280
It isn't.

10
00:00:17,280 --> 00:00:18,520
It's a fundamental shift in how we build.

11
00:00:18,520 --> 00:00:20,600
It changes identity governance and orchestration.

12
00:00:20,600 --> 00:00:22,040
It's the infrastructure layer.

13
00:00:22,040 --> 00:00:23,840
The stuff you can't see in a demo.

14
00:00:23,840 --> 00:00:26,720
Most organizations are building agents that are fragile.

15
00:00:26,720 --> 00:00:29,360
They pick products without understanding the model.

16
00:00:29,360 --> 00:00:31,040
They deploy without any governance.

17
00:00:31,040 --> 00:00:33,480
They end up with multiple agents doing the exact same work

18
00:00:33,480 --> 00:00:35,120
because they don't know what already exists.

19
00:00:35,120 --> 00:00:36,360
They're about to hit a wall.

20
00:00:36,360 --> 00:00:37,960
We need to move past product names.

21
00:00:37,960 --> 00:00:40,840
We need to look past what's in preview and what's ready now.

22
00:00:40,840 --> 00:00:44,080
We have to look at the structural reality of how agents work.

23
00:00:44,080 --> 00:00:46,320
Because that determines if your deployment scales

24
00:00:46,320 --> 00:00:47,440
or if it fails.

25
00:00:47,440 --> 00:00:50,280
Let's start with the model that makes sense of the noise.

26
00:00:50,280 --> 00:00:52,600
The four layers, how to think about agents.

27
00:00:52,600 --> 00:00:53,880
The problem is simple.

28
00:00:53,880 --> 00:00:55,320
There are too many agent products.

29
00:00:55,320 --> 00:00:57,920
You have security, co-pilot and dynamics agents.

30
00:00:57,920 --> 00:00:59,800
You have Azure agents and co-pilot studio.

31
00:00:59,800 --> 00:01:03,200
Then there's Foundry, GitHub, co-pilot and Microsoft Scout.

32
00:01:03,200 --> 00:01:06,200
Most organizations don't have a mental model for which one to use.

33
00:01:06,200 --> 00:01:08,720
They pick based on a recommendation or a random headline.

34
00:01:08,720 --> 00:01:09,840
That isn't a strategy.

35
00:01:09,840 --> 00:01:11,160
The solution is also simple.

36
00:01:11,160 --> 00:01:13,240
Stop thinking about agents as products.

37
00:01:13,240 --> 00:01:16,040
Start thinking about them as a system with four layers.

38
00:01:16,040 --> 00:01:17,400
Each layer has different needs.

39
00:01:17,400 --> 00:01:20,120
Each layer requires different people to make decisions.

40
00:01:20,120 --> 00:01:23,080
When you confuse these layers, your deployment fails.

41
00:01:23,080 --> 00:01:24,080
Experience layer.

42
00:01:24,080 --> 00:01:25,840
This is where humans talk to agents.

43
00:01:25,840 --> 00:01:27,880
It includes co-pilot and Microsoft Scout,

44
00:01:27,880 --> 00:01:29,400
which is the always on work agent.

45
00:01:29,400 --> 00:01:32,920
It includes GitHub, co-pilot and the agents built into Word or Excel.

46
00:01:32,920 --> 00:01:35,600
You also have computer using agents that click and navigate,

47
00:01:35,600 --> 00:01:36,840
just like a person does.

48
00:01:36,840 --> 00:01:38,520
This layer is about discovery.

49
00:01:38,520 --> 00:01:40,960
Most companies start here, but they also stop here.

50
00:01:40,960 --> 00:01:42,600
And that's why their setup is fragile.

51
00:01:42,600 --> 00:01:43,360
Agent layer.

52
00:01:43,360 --> 00:01:45,080
This is where specific intelligence lives.

53
00:01:45,080 --> 00:01:47,880
Security co-pilot has agents for phishing and threat intelligence.

54
00:01:47,880 --> 00:01:51,280
Dynamics 365 has agents for sales and customer intent.

55
00:01:51,280 --> 00:01:54,160
Azure co-pilot has agents for migration and troubleshooting.

56
00:01:54,160 --> 00:01:55,720
These aren't generic assistants.

57
00:01:55,720 --> 00:01:58,920
They are built for one specific job with one specific context.

58
00:01:58,920 --> 00:02:01,400
An agent that doesn't understand your domain

59
00:02:01,400 --> 00:02:03,560
is just a chatbot that's going to hallucinate.

60
00:02:03,560 --> 00:02:06,200
But once you have a dozen of these running, you hit a problem.

61
00:02:06,200 --> 00:02:09,280
You have to figure out how to control what they can actually access.

62
00:02:09,280 --> 00:02:10,000
Run time layer.

63
00:02:10,000 --> 00:02:11,480
This is where agents execute.

64
00:02:11,480 --> 00:02:14,560
It's where they keep track of their state and call other tools.

65
00:02:14,560 --> 00:02:17,400
Foundry agent service is the platform that manages this.

66
00:02:17,400 --> 00:02:21,320
You define the logic and Microsoft handles the scaling and the networking.

67
00:02:21,320 --> 00:02:24,320
The run time manages memory, including how to do things

68
00:02:24,320 --> 00:02:25,720
and what the user prefers.

69
00:02:25,720 --> 00:02:29,200
Multiple agents live in the same run time and call each other as tools.

70
00:02:29,200 --> 00:02:31,960
A coordinator agent can delegate tasks to a specialist.

71
00:02:31,960 --> 00:02:32,880
This is infrastructure.

72
00:02:32,880 --> 00:02:34,880
It isn't a feature you just bolt on at the end.

73
00:02:34,880 --> 00:02:35,800
Governance layer.

74
00:02:35,800 --> 00:02:37,680
This is the part most organizations are missing.

75
00:02:37,680 --> 00:02:40,000
It covers identity, policy and audits.

76
00:02:40,000 --> 00:02:42,600
Agent 365 is the governance plane for all of it.

77
00:02:42,600 --> 00:02:44,920
Every agent gets its own identity in Entra.

78
00:02:44,920 --> 00:02:46,720
It isn't a shared account or a human account.

79
00:02:46,720 --> 00:02:48,080
It's a unique principle.

80
00:02:48,080 --> 00:02:50,120
That means you can use least privilege access.

81
00:02:50,120 --> 00:02:51,560
You can revoke an agent's power.

82
00:02:51,560 --> 00:02:54,200
You can audit an agent just like any other system.

83
00:02:54,200 --> 00:02:57,880
Because of purview, every action an agent takes goes into an audit log.

84
00:02:57,880 --> 00:03:01,280
You can apply the same policies to agents that you apply to your users.

85
00:03:01,280 --> 00:03:03,800
Before Agent 365, these agents were invisible.

86
00:03:03,800 --> 00:03:05,720
Now they are governed like infrastructure.

87
00:03:05,720 --> 00:03:07,160
Each layer has different owners.

88
00:03:07,160 --> 00:03:08,640
They have different controls.

89
00:03:08,640 --> 00:03:11,280
Most projects fail because the team is building in one layer

90
00:03:11,280 --> 00:03:12,760
while they think they're in another.

91
00:03:12,760 --> 00:03:14,440
Or they just skip governance entirely.

92
00:03:14,440 --> 00:03:16,080
When you understand these four layers,

93
00:03:16,080 --> 00:03:17,760
you can navigate the whole ecosystem.

94
00:03:17,760 --> 00:03:19,880
It's the difference between picking products at random

95
00:03:19,880 --> 00:03:21,920
and making a real architectural decision.

96
00:03:21,920 --> 00:03:24,800
So let's look at what actually changed at Build 2026

97
00:03:24,800 --> 00:03:26,640
to make this model real.

98
00:03:26,640 --> 00:03:29,760
Build 2026, the moment agents became infrastructure.

99
00:03:29,760 --> 00:03:32,600
Everyone heard the same headlines coming out of Build 2026.

100
00:03:32,600 --> 00:03:35,760
Computer using agents when GA, Copilot Studio got a redesign.

101
00:03:35,760 --> 00:03:39,040
Voice agents arrived and teams got a fresh set of agent features.

102
00:03:39,040 --> 00:03:40,400
It is a solid product roadmap.

103
00:03:40,400 --> 00:03:43,400
It is exactly what you would expect to see at a major conference.

104
00:03:43,400 --> 00:03:45,640
But there is a quieter headline underneath that one.

105
00:03:45,640 --> 00:03:47,160
Nobody opened the show with it.

106
00:03:47,160 --> 00:03:49,040
It did not get the keynote spotlight.

107
00:03:49,040 --> 00:03:53,920
Agent 365 moved from preview to general availability on May 1, 2026

108
00:03:53,920 --> 00:03:56,600
and Foundry Agent Service hit GA back in March.

109
00:03:56,600 --> 00:03:58,760
For the first time, you have a governance plane

110
00:03:58,760 --> 00:04:01,280
and a runtime plane actually talking to each other.

111
00:04:01,280 --> 00:04:02,680
That is not just a new feature.

112
00:04:02,680 --> 00:04:04,880
That is the moment the infrastructure became real.

113
00:04:04,880 --> 00:04:07,200
Let's look at what Agent 365 actually does.

114
00:04:07,200 --> 00:04:09,280
It treats agents as first class identities.

115
00:04:09,280 --> 00:04:13,040
Not features, not capabilities bolted onto your existing systems.

116
00:04:13,040 --> 00:04:14,040
Identities.

117
00:04:14,040 --> 00:04:16,200
Every agent gets its own identity in Entra,

118
00:04:16,200 --> 00:04:20,400
which means it is not a shared service account or a human user account mapped to a bot.

119
00:04:20,400 --> 00:04:21,520
It is its own principle.

120
00:04:21,520 --> 00:04:23,240
It has its own identity in your directory,

121
00:04:23,240 --> 00:04:25,000
just like any other actor in your system.

122
00:04:25,000 --> 00:04:27,440
That changes everything about what is actually possible.

123
00:04:27,440 --> 00:04:29,640
Because once an agent has its own identity,

124
00:04:29,640 --> 00:04:32,120
it can have its own permissions and its own audit trail.

125
00:04:32,120 --> 00:04:34,440
You can apply conditional access policies to it,

126
00:04:34,440 --> 00:04:37,840
put it through regular access reviews or revoke its credentials in an instant.

127
00:04:37,840 --> 00:04:41,600
You can see exactly what it accessed when it happened and why it was there.

128
00:04:41,600 --> 00:04:43,560
The shift here is subtle but fundamental.

129
00:04:43,560 --> 00:04:45,320
You are not just using an agent anymore.

130
00:04:45,320 --> 00:04:46,320
You are hiring one.

131
00:04:46,320 --> 00:04:50,680
It comes with all the governance machinery that you would expect when hiring any other system.

132
00:04:50,680 --> 00:04:53,560
You get purview audit logs that track every move it makes,

133
00:04:53,560 --> 00:04:56,280
and entrapolices that constrain exactly what it can reach.

134
00:04:56,280 --> 00:04:57,880
It is not a feature you just turn on.

135
00:04:57,880 --> 00:04:59,920
It is a system you are now accountable for.

136
00:04:59,920 --> 00:05:02,440
Foundry agent service hit GA in March.

137
00:05:02,440 --> 00:05:04,360
And that is the runtime piece of the puzzle.

138
00:05:04,360 --> 00:05:08,280
This is the managed platform where agents actually execute and maintain their state.

139
00:05:08,280 --> 00:05:11,560
It is where they coordinate with other agents and call tools at scale.

140
00:05:11,560 --> 00:05:15,640
You get procedural memory that lets agents learn how to do things across different sessions.

141
00:05:15,640 --> 00:05:19,720
You get user memory that persists, preferences and session memory for context.

142
00:05:19,720 --> 00:05:23,000
These agents run in sandbox sessions that Microsoft manages for you.

143
00:05:23,000 --> 00:05:28,120
Scaling is handled automatically, so you define the logic while Microsoft handles the heavy lifting.

144
00:05:28,120 --> 00:05:31,560
Before Bill 2026, Foundry was just an experimental platform.

145
00:05:31,560 --> 00:05:35,960
It was interesting but it was not enterprise grade because the governance layer did not exist yet.

146
00:05:35,960 --> 00:05:40,040
You could build agents but you had no way to govern them at the organizational level.

147
00:05:40,040 --> 00:05:42,120
Now both pieces are GA and production ready.

148
00:05:42,120 --> 00:05:44,600
They are designed to work together from the ground up.

149
00:05:44,600 --> 00:05:48,360
Agent 365 discovers and governs while Foundry hosts and executes.

150
00:05:48,360 --> 00:05:51,880
One gives you identity and policy and the other gives you runtime and orchestration.

151
00:05:51,880 --> 00:05:54,120
Neither one works without the other in a real deployment.

152
00:05:54,120 --> 00:05:55,640
That is the convergence that matters.

153
00:05:55,640 --> 00:06:00,840
For the first time, organizations have the actual infrastructure required to run agents at scale with governance baked in.

154
00:06:00,840 --> 00:06:02,360
It is not bolted on after the fact.

155
00:06:02,360 --> 00:06:05,800
It is identity first, policy driven and fully auditable.

156
00:06:05,800 --> 00:06:07,480
This is why the change is structural.

157
00:06:07,480 --> 00:06:10,600
Before Bill 2026, agents were just features.

158
00:06:10,600 --> 00:06:14,520
They were advanced features but you still just built one and hoped it did not break anything.

159
00:06:14,520 --> 00:06:16,760
You had no visibility into what it was accessing.

160
00:06:16,760 --> 00:06:19,480
If it went rogue, you had no clean way to shut it down.

161
00:06:19,480 --> 00:06:24,120
After Bill 2026 agents are infrastructure, they have identities, policies and audit trails.

162
00:06:24,120 --> 00:06:27,720
They run on managed platforms that scale and coordinate with other agents.

163
00:06:27,720 --> 00:06:30,680
They are governed, like any other system in your environment.

164
00:06:30,680 --> 00:06:32,360
That is not just a product announcement.

165
00:06:32,360 --> 00:06:37,400
That is the moment when building agents stopped being experimental and started being operational.

166
00:06:37,400 --> 00:06:39,960
The four layers we talked about are no longer just concepts.

167
00:06:39,960 --> 00:06:41,640
They are an architectural reality.

168
00:06:41,640 --> 00:06:44,440
This changes how you should be thinking about your agent strategy right now.

169
00:06:44,440 --> 00:06:47,080
The question is no longer which agent product you should try.

170
00:06:47,080 --> 00:06:51,560
The real question is how you architect agents across your entire organization.

171
00:06:51,560 --> 00:06:53,960
The experience layer, where work begins.

172
00:06:53,960 --> 00:06:55,560
So we have established the model.

173
00:06:55,560 --> 00:06:57,720
Now let's look at what each layer actually does.

174
00:06:57,720 --> 00:07:01,720
We will start with the experience layer because this is where most people encounter agents.

175
00:07:01,720 --> 00:07:02,920
This is the surface.

176
00:07:02,920 --> 00:07:04,920
This is what you interact with every day.

177
00:07:04,920 --> 00:07:07,160
The experience layer is straightforward and concept.

178
00:07:07,160 --> 00:07:11,640
It is where humans interact with agents directly without needing to think about infrastructure or governance.

179
00:07:11,640 --> 00:07:13,720
It is just you and an agent getting work done.

180
00:07:13,720 --> 00:07:16,200
But even this surface level thinking has changed.

181
00:07:16,200 --> 00:07:19,320
Microsoft Scout is the clearest example of how this works now.

182
00:07:19,320 --> 00:07:21,960
Scout is positioned as an always on personal work agent.

183
00:07:21,960 --> 00:07:23,880
It is not a chatbot you ask questions,

184
00:07:23,880 --> 00:07:26,040
and it is not something you have to manually invoke.

185
00:07:26,040 --> 00:07:28,680
It is an agent that watches what you are doing across teams,

186
00:07:28,680 --> 00:07:31,160
outlook, one drive and share point.

187
00:07:31,160 --> 00:07:32,520
It acts proactively.

188
00:07:32,520 --> 00:07:35,400
You are in a meeting and Scout monitors the conversation.

189
00:07:35,400 --> 00:07:38,040
You are preparing a document and Scout watches the context.

190
00:07:38,040 --> 00:07:39,160
Without being asked,

191
00:07:39,160 --> 00:07:42,920
it surfaces relevant information and flags decisions you need to make.

192
00:07:42,920 --> 00:07:44,760
It does not wait for you to formulate a question

193
00:07:44,760 --> 00:07:48,120
because it understands your context and anticipates what you need

194
00:07:48,120 --> 00:07:50,760
compared that to traditional co-pilot, which is reactive.

195
00:07:50,760 --> 00:07:54,040
You open word and type a prompt to write an introduction to a report.

196
00:07:54,040 --> 00:07:55,880
Co-pilot responds, and it is helpful,

197
00:07:55,880 --> 00:07:57,400
but it is still a Q&A model.

198
00:07:57,400 --> 00:07:58,600
You ask and it answers.

199
00:07:58,600 --> 00:07:59,960
Scout flips that entirely.

200
00:07:59,960 --> 00:08:01,320
It observes and then it acts.

201
00:08:01,320 --> 00:08:04,200
That is a fundamentally different way to interact with software.

202
00:08:04,200 --> 00:08:06,920
GitHub co-pilot CLI brings that same agent concept

203
00:08:06,920 --> 00:08:09,160
to developers in a completely different environment.

204
00:08:09,160 --> 00:08:12,120
You are in a terminal and you need help with a Git workflow

205
00:08:12,120 --> 00:08:13,880
or a command you cannot remember.

206
00:08:13,880 --> 00:08:16,200
Instead of switching to a browser to look at documentation,

207
00:08:16,200 --> 00:08:18,200
the agent operates right there in your terminal.

208
00:08:18,200 --> 00:08:21,400
It is the same principle as Scout, just on a different surface.

209
00:08:21,400 --> 00:08:24,440
Then you have agents embedded in the applications themselves.

210
00:08:24,440 --> 00:08:26,840
Word has an agent that understands the tone and audience

211
00:08:26,840 --> 00:08:28,200
of the document you are writing.

212
00:08:28,200 --> 00:08:30,280
Excel has an agent that knows your data structure

213
00:08:30,280 --> 00:08:31,720
and helps you analyze it.

214
00:08:31,720 --> 00:08:35,480
PowerPoint has an agent that understands the story you are telling with your slides.

215
00:08:35,480 --> 00:08:37,320
These are not generic assistance.

216
00:08:37,320 --> 00:08:40,200
They understand the specific context of the work happening

217
00:08:40,200 --> 00:08:41,400
in that specific application.

218
00:08:41,400 --> 00:08:44,760
Co-pilot Studio sits somewhere different on this layer.

219
00:08:44,760 --> 00:08:48,920
It is the tool that lets teams build custom agents without writing any code.

220
00:08:48,920 --> 00:08:51,800
You use a low-code interface to define the behavior you want

221
00:08:51,800 --> 00:08:53,080
and connect your knowledge sources.

222
00:08:53,080 --> 00:08:54,600
You test it and then you launch it.

223
00:08:54,600 --> 00:08:58,600
Most organizations start here because the bar to entry is genuinely low.

224
00:08:58,600 --> 00:09:01,320
Computer using agents take a different approach entirely.

225
00:09:01,320 --> 00:09:03,880
Instead of understanding natural language and calling APIs,

226
00:09:03,880 --> 00:09:05,480
these agents see your screen.

227
00:09:05,480 --> 00:09:07,080
They watch your UI, click like you do,

228
00:09:07,080 --> 00:09:09,080
and navigate exactly like a human would.

229
00:09:09,080 --> 00:09:12,360
This matters for legacy systems that do not have clean integrations.

230
00:09:12,360 --> 00:09:15,320
A computer using agent can work with any system a human can work with.

231
00:09:15,320 --> 00:09:17,960
Here is the pattern you see across the experience layer.

232
00:09:17,960 --> 00:09:21,320
It is all about accessibility and meeting people where they already work.

233
00:09:21,320 --> 00:09:24,120
Whether it is in teams, a terminal, or a calendar,

234
00:09:24,120 --> 00:09:25,800
the infrastructure is not the focus.

235
00:09:25,800 --> 00:09:29,080
The experience is, you want to know how quickly you can get value

236
00:09:29,080 --> 00:09:31,960
and how well the agent fits into your existing workflow.

237
00:09:31,960 --> 00:09:33,560
Most organizations start right here.

238
00:09:33,560 --> 00:09:36,200
They try Scout, build something in co-pilot studio,

239
00:09:36,200 --> 00:09:38,200
and enable co-pilot in their apps.

240
00:09:38,200 --> 00:09:40,200
They see the value and they think they are done.

241
00:09:40,200 --> 00:09:42,120
And that is where the fragility comes in.

242
00:09:42,120 --> 00:09:44,200
Because the experience layer is only one layer.

243
00:09:44,200 --> 00:09:46,520
It is the visible part, but underneath,

244
00:09:46,520 --> 00:09:50,040
there is a whole architecture that determines whether this actually scales.

245
00:09:50,040 --> 00:09:52,360
It determines whether your data stays secure

246
00:09:52,360 --> 00:09:56,280
and whether you can actually manage these agents according to your company policies.

247
00:09:56,280 --> 00:09:59,320
The agent layer, domain-specific intelligence.

248
00:09:59,320 --> 00:10:02,120
Under the surface experience, something else is operating.

249
00:10:02,120 --> 00:10:04,600
The agents themselves, and this is where the model matters.

250
00:10:04,600 --> 00:10:08,360
Because an agent built for security is fundamentally different from an agent built for sales.

251
00:10:08,360 --> 00:10:11,320
They have different training, different constraints, different context.

252
00:10:11,320 --> 00:10:12,440
They serve a different purpose.

253
00:10:12,440 --> 00:10:13,640
That is the agent layer.

254
00:10:13,640 --> 00:10:16,840
The agent layer is where domain-specific intelligence actually lives.

255
00:10:16,840 --> 00:10:18,280
These are not generic assistants.

256
00:10:18,280 --> 00:10:20,280
They are not chatbots that sort of help with anything.

257
00:10:20,280 --> 00:10:24,600
These are agents built for a specific job in a specific context with specific constraints.

258
00:10:24,600 --> 00:10:25,880
They understand deeply.

259
00:10:25,880 --> 00:10:28,360
Take security co-pilot agents as an example.

260
00:10:28,360 --> 00:10:31,400
These are not general-purpose assistants that happen to work on security.

261
00:10:31,400 --> 00:10:34,360
They were built from the ground up to understand security signals,

262
00:10:34,360 --> 00:10:36,760
threat patterns, and remediation workflows.

263
00:10:36,760 --> 00:10:39,960
There is a phishing triage agent that analyzes suspicious messages.

264
00:10:39,960 --> 00:10:42,200
It separates likely phishing from legitimate mail

265
00:10:42,200 --> 00:10:44,840
and helps analysts prioritize what to investigate.

266
00:10:44,840 --> 00:10:47,480
That is not a generic language model answering questions.

267
00:10:47,480 --> 00:10:51,960
It is a specialist agent that understands email signals the way security teams understand them.

268
00:10:51,960 --> 00:10:54,040
Alert triage agents work the same way.

269
00:10:54,040 --> 00:10:56,440
Security teams get flooded with alerts every day.

270
00:10:56,440 --> 00:10:58,040
Most are noise, but some are signals.

271
00:10:58,040 --> 00:11:00,760
The alert triage agent learns to distinguish between them

272
00:11:00,760 --> 00:11:02,520
because it understands context.

273
00:11:02,520 --> 00:11:05,480
It knows which alerts cluster together to suggest a larger incident

274
00:11:05,480 --> 00:11:07,960
and surfaces what actually requires human attention.

275
00:11:07,960 --> 00:11:09,800
It was built for that specific problem.

276
00:11:09,800 --> 00:11:12,680
Conditional access agents, vulnerability remediation agents,

277
00:11:12,680 --> 00:11:14,840
and threat intelligence agents are all specialized.

278
00:11:14,840 --> 00:11:17,960
Each one has been trained on domain-specific data and workflows.

279
00:11:17,960 --> 00:11:19,560
Each one has constraints built in

280
00:11:19,560 --> 00:11:22,920
because security people understand what bad output looks like

281
00:11:22,920 --> 00:11:24,040
and what could cause harm.

282
00:11:24,040 --> 00:11:25,800
Then look at Dynamics 365 agents.

283
00:11:25,800 --> 00:11:27,320
This is a completely different domain.

284
00:11:27,320 --> 00:11:29,640
The sales qualification agent is not a chatbot.

285
00:11:29,640 --> 00:11:31,720
It was built to understand lead scoring criteria,

286
00:11:31,720 --> 00:11:33,640
customer fit, and pipeline progression.

287
00:11:33,640 --> 00:11:36,840
It assesses inbound leads against what your sales team actually cares about

288
00:11:36,840 --> 00:11:39,240
and surfaces high-probability opportunities

289
00:11:39,240 --> 00:11:40,840
instead of everything that comes in.

290
00:11:40,840 --> 00:11:43,560
The account reconciliation agent solves a different problem.

291
00:11:43,560 --> 00:11:46,120
Finance teams spend enormous amounts of time matching records

292
00:11:46,120 --> 00:11:49,080
like invoices to payments or purchase orders to receipts.

293
00:11:49,080 --> 00:11:50,920
The agent understands matching logic

294
00:11:50,920 --> 00:11:52,520
and knows what constitutes a match.

295
00:11:52,520 --> 00:11:55,000
It flags exceptions and reduces the hours human spend

296
00:11:55,000 --> 00:11:56,360
on purely mechanical work.

297
00:11:56,360 --> 00:11:59,000
The customer intent agent works across multiple channels.

298
00:11:59,000 --> 00:12:01,800
Sales and service teams have interactions through email, calls,

299
00:12:01,800 --> 00:12:03,480
web forms, and social media.

300
00:12:03,480 --> 00:12:05,720
The agent analyzes those signals to understand

301
00:12:05,720 --> 00:12:07,320
what the customer actually wants

302
00:12:07,320 --> 00:12:09,560
underneath what they explicitly asked for.

303
00:12:09,560 --> 00:12:11,240
This helps teams respond with better timing

304
00:12:11,240 --> 00:12:12,360
and better solutions.

305
00:12:12,360 --> 00:12:14,360
Supplyer communications and field service

306
00:12:14,360 --> 00:12:16,840
each have their own agents built for specific domains

307
00:12:16,840 --> 00:12:18,360
with specific workflows.

308
00:12:18,360 --> 00:12:20,120
Azure brings its own set to the table.

309
00:12:20,120 --> 00:12:23,800
There are migration agents for assessing and planning complex Azure migrations

310
00:12:23,800 --> 00:12:26,200
and deployment agents that guide implementation

311
00:12:26,200 --> 00:12:27,960
once the architecture is decided.

312
00:12:27,960 --> 00:12:30,200
Optimization agents run on existing environments

313
00:12:30,200 --> 00:12:32,600
to improve cost, performance, and utilization.

314
00:12:32,600 --> 00:12:35,240
Observability agents connect disparate telemetry signals

315
00:12:35,240 --> 00:12:37,880
to surface what is actually happening in your infrastructure.

316
00:12:37,880 --> 00:12:40,120
Resiliency agents strengthen fault tolerance

317
00:12:40,120 --> 00:12:41,320
before failure happens.

318
00:12:41,320 --> 00:12:44,680
And troubleshooting agents diagnose problems faster than humans can.

319
00:12:44,680 --> 00:12:46,040
And then there is the open category.

320
00:12:46,040 --> 00:12:48,440
Copilot Studio lets teams build custom agents

321
00:12:48,440 --> 00:12:50,200
for business specific workflows

322
00:12:50,200 --> 00:12:52,280
that do not fit any pre-built pattern.

323
00:12:52,280 --> 00:12:54,040
These are processes unique to your industry

324
00:12:54,040 --> 00:12:56,280
or workflows specific to your organization.

325
00:12:56,280 --> 00:12:58,040
And those agents are built by business teams

326
00:12:58,040 --> 00:12:59,800
who understand that context deeply.

327
00:12:59,800 --> 00:13:02,120
The key distinction is that these are not generic assistance

328
00:13:02,120 --> 00:13:04,040
with domain knowledge bolted on.

329
00:13:04,040 --> 00:13:06,200
They were built from the ground up for a specific job.

330
00:13:06,200 --> 00:13:08,280
They have constraints, they have specialized training.

331
00:13:08,280 --> 00:13:10,040
They understand the specific signals

332
00:13:10,040 --> 00:13:11,400
that matter in that domain.

333
00:13:11,400 --> 00:13:14,040
This means an agent that does not understand your domain

334
00:13:14,040 --> 00:13:16,200
is just a chatbot with hallucination risk.

335
00:13:16,200 --> 00:13:18,360
A security agent applied to sales is useless.

336
00:13:18,360 --> 00:13:20,760
A sales agent applied to operations is worse than useless

337
00:13:20,760 --> 00:13:22,680
because it will be confidently wrong.

338
00:13:22,680 --> 00:13:24,520
But this creates a governance problem.

339
00:13:24,520 --> 00:13:27,960
Once you have dozens of specialized agents running across your organization,

340
00:13:27,960 --> 00:13:29,640
you have to control what they access.

341
00:13:29,640 --> 00:13:32,520
Some are pre-built, some are custom, and some are from vendors.

342
00:13:32,520 --> 00:13:34,280
You need to keep them aligned with policy

343
00:13:34,280 --> 00:13:35,960
and you need to know they even exist.

344
00:13:35,960 --> 00:13:39,240
That is where the runtime and governance layers become critical.

345
00:13:39,240 --> 00:13:42,200
The runtime layer, where agents actually execute,

346
00:13:42,200 --> 00:13:44,520
the infrastructure where agents actually run

347
00:13:44,520 --> 00:13:45,880
is a different problem entirely.

348
00:13:45,880 --> 00:13:48,040
Once you have agents with specialized capabilities,

349
00:13:48,040 --> 00:13:49,320
they need somewhere to live.

350
00:13:49,320 --> 00:13:52,680
They need a place to maintain state, call tools and connect to data.

351
00:13:52,680 --> 00:13:55,240
They need to coordinate with other agents doing related work.

352
00:13:55,240 --> 00:13:56,680
That is the runtime layer.

353
00:13:56,680 --> 00:13:59,000
Foundry agent service is the manage platform for this.

354
00:13:59,000 --> 00:14:00,120
Think of it like this.

355
00:14:00,120 --> 00:14:02,440
You write the agent logic and define what it should do.

356
00:14:02,440 --> 00:14:04,680
You specify what tools it needs access to

357
00:14:04,680 --> 00:14:06,040
and then you hand it to Foundry.

358
00:14:06,040 --> 00:14:07,800
Foundry handles everything underneath,

359
00:14:07,800 --> 00:14:10,840
including scaling, isolation, networking, and identity.

360
00:14:10,840 --> 00:14:13,240
It manages session management and state persistence

361
00:14:13,240 --> 00:14:16,760
so you can focus on the logic while Microsoft manages the infrastructure.

362
00:14:16,760 --> 00:14:18,920
That distinction matters more than it sounds.

363
00:14:18,920 --> 00:14:21,400
Building agents yourself means managing containers,

364
00:14:21,400 --> 00:14:23,640
scaling groups, and databases for state.

365
00:14:23,640 --> 00:14:25,720
You have to manage memory and networking.

366
00:14:25,720 --> 00:14:28,680
Building agents on Foundry means you are not managing any of that

367
00:14:28,680 --> 00:14:30,120
because the platform does it for you.

368
00:14:30,120 --> 00:14:33,000
You define the agent and deploy it, and then it runs.

369
00:14:33,000 --> 00:14:35,560
This is what managed actually means in practice.

370
00:14:35,560 --> 00:14:39,160
Agents run in hosted sessions where each session gets its own sandbox.

371
00:14:39,160 --> 00:14:41,800
The session has file system access and persistent state

372
00:14:41,800 --> 00:14:43,560
that survives across interactions.

373
00:14:43,560 --> 00:14:46,840
When a user talks to an agent that conversation maintains context,

374
00:14:46,840 --> 00:14:50,760
when the agent needs to do work, it has a place to store intermediate results.

375
00:14:50,760 --> 00:14:52,600
When another agent needs to call this agent,

376
00:14:52,600 --> 00:14:54,600
the call goes through a standard protocol.

377
00:14:54,600 --> 00:14:57,240
Microsoft handles the isolation and the scaling.

378
00:14:57,240 --> 00:15:00,040
If traffic spikes, the platform scales automatically,

379
00:15:00,040 --> 00:15:02,680
and if one agent fails, it does not cascade.

380
00:15:02,680 --> 00:15:05,560
Memory is built into the runtime instead of being bolted on.

381
00:15:05,560 --> 00:15:07,800
This is where the model shows its maturity.

382
00:15:07,800 --> 00:15:10,520
Procedural memory means agents remember how to do things.

383
00:15:10,520 --> 00:15:13,400
Once they have executed a workflow, they remember the steps,

384
00:15:13,400 --> 00:15:16,040
which makes them faster and more accurate the second time.

385
00:15:16,040 --> 00:15:18,040
User memory persists across sessions,

386
00:15:18,040 --> 00:15:20,280
so an agent learns your preferences, your constraints,

387
00:15:20,280 --> 00:15:21,480
and your typical patterns.

388
00:15:21,480 --> 00:15:22,440
It carries that forward.

389
00:15:22,440 --> 00:15:25,480
Session memory holds context for the current conversation,

390
00:15:25,480 --> 00:15:27,320
but only the current conversation.

391
00:15:27,320 --> 00:15:29,720
Once the session ends, session memory goes away,

392
00:15:29,720 --> 00:15:31,880
but procedural and user memory stick around.

393
00:15:31,880 --> 00:15:34,120
Multi-agent orchestration happens in the runtime.

394
00:15:34,120 --> 00:15:36,440
One agent can call another agent as a tool.

395
00:15:36,440 --> 00:15:39,560
You have a coordinator agent that breaks down a complex task

396
00:15:39,560 --> 00:15:41,640
and calls a data retrieval agent.

397
00:15:41,640 --> 00:15:43,240
That agent goes to fetch information

398
00:15:43,240 --> 00:15:44,920
and then calls an analysis agent.

399
00:15:44,920 --> 00:15:46,840
That agent processes what was retrieved

400
00:15:46,840 --> 00:15:49,320
and calls a communication agent to draft a response.

401
00:15:49,320 --> 00:15:51,240
The coordinator stitches it all together

402
00:15:51,240 --> 00:15:53,400
while the runtime handles the choreography.

403
00:15:53,400 --> 00:15:54,920
Each agent knows what tool it is calling

404
00:15:54,920 --> 00:15:56,200
and what response to expect.

405
00:15:56,200 --> 00:15:59,480
If one agent fails, the coordinator can retry or escalate.

406
00:15:59,480 --> 00:16:02,280
The agent to agent protocol standardizes this process.

407
00:16:02,280 --> 00:16:04,200
Agents do not just work within foundry.

408
00:16:04,200 --> 00:16:06,120
They can call agents built on other frameworks

409
00:16:06,120 --> 00:16:08,680
as long as they speak the A2A protocol.

410
00:16:08,680 --> 00:16:11,880
Agents built on semantic kernel can call agents built on auto-gen.

411
00:16:11,880 --> 00:16:13,800
Custom agents can call pre-built agents.

412
00:16:13,800 --> 00:16:16,440
They do not all have to run on the same platform

413
00:16:16,440 --> 00:16:18,280
as long as they understand the protocol.

414
00:16:18,280 --> 00:16:19,960
Toolboxes are the discovery mechanism.

415
00:16:19,960 --> 00:16:22,680
Instead of hard-coding every tool an agent might need,

416
00:16:22,680 --> 00:16:25,400
toolboxes let agents discover tools at runtime.

417
00:16:25,400 --> 00:16:28,600
A toolbox is a collection of tools, integrations, and data sources.

418
00:16:28,600 --> 00:16:30,680
An agent looks at the task it is facing

419
00:16:30,680 --> 00:16:32,840
and checks what is available in the toolbox.

420
00:16:32,840 --> 00:16:34,920
It selects what is relevant and uses it.

421
00:16:34,920 --> 00:16:37,080
If new tools get added to the toolbox,

422
00:16:37,080 --> 00:16:39,080
every agent automatically has access

423
00:16:39,080 --> 00:16:41,720
without any redeployment or reconfiguration.

424
00:16:41,720 --> 00:16:44,120
Foundry IQ is the knowledge layer underneath.

425
00:16:44,120 --> 00:16:47,160
It unifies retrieval across documents, data warehouses,

426
00:16:47,160 --> 00:16:49,160
web content, and enterprise sources.

427
00:16:49,160 --> 00:16:51,240
An agent that needs to answer a question

428
00:16:51,240 --> 00:16:53,560
does not have to know where the answer lives.

429
00:16:53,560 --> 00:16:55,800
It asks Foundry IQ and the platform finds it.

430
00:16:55,800 --> 00:16:57,240
It returns relevant context

431
00:16:57,240 --> 00:17:00,360
and the agent uses that context to construct a response.

432
00:17:00,360 --> 00:17:02,040
There are no custom rag pipelines

433
00:17:02,040 --> 00:17:04,280
to build a no-vector databases to manage.

434
00:17:04,280 --> 00:17:06,120
Foundry IQ handles the orchestration

435
00:17:06,120 --> 00:17:07,960
of multiple retrieval sources.

436
00:17:07,960 --> 00:17:09,880
The reason runtime matters is simple.

437
00:17:09,880 --> 00:17:11,240
Without a managed runtime,

438
00:17:11,240 --> 00:17:14,040
you are building agents that run in your own infrastructure.

439
00:17:14,040 --> 00:17:15,480
You are responsible for scaling,

440
00:17:15,480 --> 00:17:17,480
state management, and keeping them secure.

441
00:17:17,480 --> 00:17:20,600
You have to make sure they do not break other systems.

442
00:17:20,600 --> 00:17:22,120
With Foundry, you are building agents

443
00:17:22,120 --> 00:17:23,960
that scale automatically and maintain

444
00:17:23,960 --> 00:17:25,160
state without your involvement.

445
00:17:25,160 --> 00:17:26,520
They coordinate with other agents

446
00:17:26,520 --> 00:17:28,760
without custom orchestration and access tools

447
00:17:28,760 --> 00:17:30,760
without hard-coded integrations.

448
00:17:30,760 --> 00:17:31,960
The runtime is what lets you move

449
00:17:31,960 --> 00:17:35,080
from experimental agents to operational agents at scale.

450
00:17:35,080 --> 00:17:37,480
The governance layer sits on top and controls all of this

451
00:17:37,480 --> 00:17:39,960
but the runtime layer is what makes it actually work.

452
00:17:39,960 --> 00:17:42,200
The governance layer, the missing piece,

453
00:17:42,200 --> 00:17:44,440
the runtime layer gives you the basics.

454
00:17:44,440 --> 00:17:47,000
Hosting, memory, scaling.

455
00:17:47,000 --> 00:17:49,480
But infrastructure without control is just chaos.

456
00:17:49,480 --> 00:17:51,480
That is where the governance layer comes in.

457
00:17:51,480 --> 00:17:54,920
Before agent 365 agents were essentially invisible.

458
00:17:54,920 --> 00:17:57,080
You would build an agent in Co-Pilot Studio

459
00:17:57,080 --> 00:17:58,600
and it would just run somewhere.

460
00:17:58,600 --> 00:17:59,960
It might be in the power platform

461
00:17:59,960 --> 00:18:01,240
or a custom environment

462
00:18:01,240 --> 00:18:03,240
or even on a laptop running a local script.

463
00:18:03,240 --> 00:18:04,440
Nobody knew it was there.

464
00:18:04,440 --> 00:18:05,400
It did not know.

465
00:18:05,400 --> 00:18:06,520
Security did not know.

466
00:18:06,520 --> 00:18:08,520
The organization had no idea what was operating,

467
00:18:08,520 --> 00:18:09,880
what data it was touching

468
00:18:09,880 --> 00:18:11,560
or who was even responsible for it.

469
00:18:11,560 --> 00:18:14,840
That invisibility was fine when agents were just experiments.

470
00:18:14,840 --> 00:18:18,120
It became dangerous the moment they started calling real APIs

471
00:18:18,120 --> 00:18:19,800
and making decisions at scale.

472
00:18:19,800 --> 00:18:21,160
The governance layer changes that.

473
00:18:21,160 --> 00:18:22,600
It makes agents visible.

474
00:18:22,600 --> 00:18:24,840
More importantly, it makes them manageable.

475
00:18:24,840 --> 00:18:27,320
Agent 365 acts as a central registry.

476
00:18:27,320 --> 00:18:29,400
It discovers every agent in your tenant

477
00:18:29,400 --> 00:18:31,320
not just the ones you officially deployed.

478
00:18:31,320 --> 00:18:33,640
It finds the agents in power platform environments

479
00:18:33,640 --> 00:18:35,320
that have not been touched in six months

480
00:18:35,320 --> 00:18:37,720
and it surfaces the Co-Pilot Studio bot someone built

481
00:18:37,720 --> 00:18:38,760
in a private workspace.

482
00:18:38,760 --> 00:18:41,480
It discovers shadow agents that nobody told it about.

483
00:18:41,480 --> 00:18:43,880
For the first time, you can actually see what is running.

484
00:18:43,880 --> 00:18:45,240
But discovery is only the start.

485
00:18:45,240 --> 00:18:47,000
Once you see an agent, you have to control it.

486
00:18:47,000 --> 00:18:49,400
And that is where EntraAgentID changes the model.

487
00:18:49,400 --> 00:18:52,040
Every agent gets its own identity in your EntraDirectory.

488
00:18:52,040 --> 00:18:53,640
This is not a shared service account

489
00:18:53,640 --> 00:18:55,400
or a human account mapped to a bot.

490
00:18:55,400 --> 00:18:56,760
It is its own principle.

491
00:18:56,760 --> 00:18:58,680
When an agent acts, it acts as itself.

492
00:18:58,680 --> 00:19:00,440
It has its own permissions, its own audit trail

493
00:19:00,440 --> 00:19:01,400
and its own life cycle.

494
00:19:01,400 --> 00:19:03,000
That sounds like a minor detail.

495
00:19:03,000 --> 00:19:04,600
It is not. It is foundational.

496
00:19:04,600 --> 00:19:08,040
Because an agent with its own identity

497
00:19:08,040 --> 00:19:09,640
can follow the rule of least privilege.

498
00:19:09,640 --> 00:19:11,080
If an agent only needs to read

499
00:19:11,080 --> 00:19:12,840
from one specific SharePoint library,

500
00:19:12,840 --> 00:19:15,640
it gets read-only access to that library and nothing else.

501
00:19:15,640 --> 00:19:18,280
It does not get broad access to all of SharePoint.

502
00:19:18,280 --> 00:19:19,640
It does not use a service account

503
00:19:19,640 --> 00:19:21,400
that can read every file in the company.

504
00:19:21,400 --> 00:19:23,320
It gets exactly what it needs to do the job.

505
00:19:23,320 --> 00:19:25,800
That is the opposite of how agents usually work today.

506
00:19:25,800 --> 00:19:27,800
Right now, they either have no identity at all

507
00:19:27,800 --> 00:19:29,880
or they run under a service account with permissions

508
00:19:29,880 --> 00:19:32,440
that were set once and never checked again.

509
00:19:32,440 --> 00:19:35,720
With EntraEgentID, every agent is secure by design.

510
00:19:35,720 --> 00:19:37,720
Because every agent has its own identity,

511
00:19:37,720 --> 00:19:40,440
you can scope that identity with total precision.

512
00:19:40,440 --> 00:19:42,520
If an agent goes rogue or gets compromised,

513
00:19:42,520 --> 00:19:44,360
you simply revoke its identity.

514
00:19:44,360 --> 00:19:45,560
It stops immediately.

515
00:19:45,560 --> 00:19:48,280
It does not wait for a manual shutdown or a service restart.

516
00:19:48,280 --> 00:19:50,280
It dies the moment the identity is gone.

517
00:19:50,280 --> 00:19:53,000
Every action an agent takes is logged under its own name.

518
00:19:53,000 --> 00:19:54,360
Because of Pervue integration,

519
00:19:54,360 --> 00:19:56,280
that audit trail flows into the same logs

520
00:19:56,280 --> 00:19:58,040
that track everything else in your environment.

521
00:19:58,040 --> 00:19:59,640
You can see what every agent accessed

522
00:19:59,640 --> 00:20:00,840
and exactly what it did.

523
00:20:00,840 --> 00:20:03,080
For compliance or for figuring out what went wrong,

524
00:20:03,080 --> 00:20:04,760
the trail is finally complete.

525
00:20:04,760 --> 00:20:07,320
The same conditional access policies you apply to humans

526
00:20:07,320 --> 00:20:08,760
now apply to agents.

527
00:20:08,760 --> 00:20:10,600
If your policy says sensitive data

528
00:20:10,600 --> 00:20:12,600
must only come from a managed device

529
00:20:12,600 --> 00:20:14,520
that rule applies to your agents too.

530
00:20:14,520 --> 00:20:16,680
If you restrict access based on risk or location,

531
00:20:16,680 --> 00:20:18,520
agents have to follow those rules.

532
00:20:18,520 --> 00:20:20,440
You are no longer managing agent security

533
00:20:20,440 --> 00:20:21,400
in a separate silo.

534
00:20:21,400 --> 00:20:22,520
It is all one system.

535
00:20:22,520 --> 00:20:24,040
The agent control specification

536
00:20:24,040 --> 00:20:26,280
lets you define what agents are allowed to do.

537
00:20:26,280 --> 00:20:27,800
This is different from access control.

538
00:20:27,800 --> 00:20:28,680
It is about actions.

539
00:20:28,680 --> 00:20:31,480
Maybe an agent can read documents but cannot modify them.

540
00:20:31,480 --> 00:20:33,640
Maybe it can send a message but cannot delete one.

541
00:20:33,640 --> 00:20:35,080
You define the boundaries.

542
00:20:35,080 --> 00:20:36,680
And the runtime enforces them.

543
00:20:36,680 --> 00:20:37,880
This is the structural shift.

544
00:20:37,880 --> 00:20:41,400
Before agent 365 agents were invisible and uncontrolled.

545
00:20:41,400 --> 00:20:43,400
Now, they are governed like any other system.

546
00:20:43,400 --> 00:20:44,600
They have identity.

547
00:20:44,600 --> 00:20:45,560
They have audits.

548
00:20:45,560 --> 00:20:46,760
They have a life cycle.

549
00:20:46,760 --> 00:20:48,120
Governance is not an add-on.

550
00:20:48,120 --> 00:20:49,560
It is infrastructure.

551
00:20:49,560 --> 00:20:50,840
The decision framework,

552
00:20:50,840 --> 00:20:52,520
Azure and GitHub agents.

553
00:20:52,520 --> 00:20:53,960
We have covered the layers.

554
00:20:53,960 --> 00:20:55,080
You understand the model.

555
00:20:55,080 --> 00:20:58,040
But the model does not tell you which agent to actually use.

556
00:20:58,040 --> 00:20:59,640
That requires a decision framework.

557
00:20:59,640 --> 00:21:01,320
And it starts with a simple question.

558
00:21:01,320 --> 00:21:03,560
Is the work you are doing focused on infrastructure

559
00:21:03,560 --> 00:21:04,680
or on developers?

560
00:21:04,680 --> 00:21:07,080
If the work is about deployment, operations,

561
00:21:07,080 --> 00:21:08,680
or cloud optimization,

562
00:21:08,680 --> 00:21:10,280
you are looking at Azure agents.

563
00:21:10,280 --> 00:21:11,800
These are specialized for the cloud.

564
00:21:11,800 --> 00:21:13,160
They understand the specific problems

565
00:21:13,160 --> 00:21:15,240
that operations teams deal with every day.

566
00:21:15,240 --> 00:21:16,600
Take the migration agent.

567
00:21:16,600 --> 00:21:18,520
You use this when you are planning a move to Azure

568
00:21:18,520 --> 00:21:19,960
with complex dependencies.

569
00:21:19,960 --> 00:21:21,480
Your environment likely has systems

570
00:21:21,480 --> 00:21:23,720
that talk to each other in ways you do not fully see.

571
00:21:23,720 --> 00:21:26,200
The migration agent understands those relationships.

572
00:21:26,200 --> 00:21:27,960
It maps what is connected to what

573
00:21:27,960 --> 00:21:29,800
and finds the dependencies you might miss

574
00:21:29,800 --> 00:21:31,000
during a manual check.

575
00:21:31,000 --> 00:21:32,920
It surfaces the risks of moving system A

576
00:21:32,920 --> 00:21:34,040
before system B.

577
00:21:34,040 --> 00:21:35,560
This speeds up the assessment phase

578
00:21:35,560 --> 00:21:37,000
where most companies waste months

579
00:21:37,000 --> 00:21:38,280
just documenting what they own.

580
00:21:38,280 --> 00:21:40,440
The deployment agent is for a different stage.

581
00:21:40,440 --> 00:21:43,160
Use this when the architecture decisions are already made.

582
00:21:43,160 --> 00:21:44,520
You know where the workloads are going

583
00:21:44,520 --> 00:21:45,880
and what services you need.

584
00:21:45,880 --> 00:21:47,560
Now you need a guide for implementation.

585
00:21:47,560 --> 00:21:49,000
How do you deploy this at scale?

586
00:21:49,000 --> 00:21:50,040
How do you make it repeatable

587
00:21:50,040 --> 00:21:52,120
to the fifth deployment is as clean as the first?

588
00:21:52,120 --> 00:21:53,640
It helps you build a process

589
00:21:53,640 --> 00:21:55,560
so someone else can deploy the same setup

590
00:21:55,560 --> 00:21:57,560
without needing the knowledge inside your head.

591
00:21:57,560 --> 00:21:59,320
The optimization agent runs on environments

592
00:21:59,320 --> 00:22:00,520
that are already live.

593
00:22:00,520 --> 00:22:02,280
Use it when you are past the setup phase

594
00:22:02,280 --> 00:22:03,720
and into a steady state.

595
00:22:03,720 --> 00:22:05,160
Your infrastructure is running

596
00:22:05,160 --> 00:22:07,000
but you might be spending too much on compute

597
00:22:07,000 --> 00:22:08,600
or leaving performance on the table.

598
00:22:08,600 --> 00:22:11,000
This agent analyzes what you are actually using

599
00:22:11,000 --> 00:22:12,600
versus what you are paying for.

600
00:22:12,600 --> 00:22:14,360
It finds right sizing opportunities

601
00:22:14,360 --> 00:22:15,720
and unused resources.

602
00:22:15,720 --> 00:22:17,400
It is not about changing the architecture.

603
00:22:17,400 --> 00:22:20,040
It is about making what you have run more efficiently.

604
00:22:20,040 --> 00:22:22,760
The observability agent is for when you have plenty of data

605
00:22:22,760 --> 00:22:24,760
but the signal is buried in noise.

606
00:22:24,760 --> 00:22:27,000
Your systems are producing logs, metrics

607
00:22:27,000 --> 00:22:27,880
and traces.

608
00:22:27,880 --> 00:22:30,040
Connecting those signals into a real understanding

609
00:22:30,040 --> 00:22:32,200
of what is happening is the hard part.

610
00:22:32,200 --> 00:22:34,840
The observability agent looks across all your data sources

611
00:22:34,840 --> 00:22:36,120
to correlate events.

612
00:22:36,120 --> 00:22:38,520
It helps you understand not just that something failed

613
00:22:38,520 --> 00:22:39,480
but why it happened.

614
00:22:39,480 --> 00:22:42,360
The resiliency agent works before things break.

615
00:22:42,360 --> 00:22:45,000
Use this when you want to strengthen your fall tolerance.

616
00:22:45,000 --> 00:22:46,760
What happens if a database goes down?

617
00:22:46,760 --> 00:22:48,680
What happens if a network connection fails?

618
00:22:48,680 --> 00:22:51,320
The resiliency agent looks for weak points in your architecture.

619
00:22:51,320 --> 00:22:52,680
It suggests improvements

620
00:22:52,680 --> 00:22:54,280
so you can build in redundancy

621
00:22:54,280 --> 00:22:55,320
where it actually matters.

622
00:22:56,120 --> 00:22:58,120
The troubleshooting agent is for when things go wrong.

623
00:22:58,120 --> 00:23:00,120
Something broke and you need a diagnosis fast.

624
00:23:00,120 --> 00:23:02,760
The problem could be in the network, the app or the database.

625
00:23:02,760 --> 00:23:05,080
This agent pulls signals from every source at once.

626
00:23:05,080 --> 00:23:06,920
It diagnosis faster than a human

627
00:23:06,920 --> 00:23:09,080
because it does not have to switch between different tools

628
00:23:09,080 --> 00:23:09,880
and documentation.

629
00:23:09,880 --> 00:23:11,560
It sees everything at the same time.

630
00:23:11,560 --> 00:23:14,040
GitHub Copilot CLI is the version for developers.

631
00:23:14,040 --> 00:23:17,320
Use this when teams need help without leaving their workflow.

632
00:23:17,320 --> 00:23:19,160
They are in the terminal and they need a command

633
00:23:19,160 --> 00:23:20,760
or help with an error message.

634
00:23:20,760 --> 00:23:22,280
Instead of switching to a browser,

635
00:23:22,280 --> 00:23:24,280
the agent operates right there in the environment

636
00:23:24,280 --> 00:23:25,560
where the work is happening.

637
00:23:25,560 --> 00:23:28,280
The modernization agent helps teams update old code

638
00:23:28,280 --> 00:23:29,720
without starting from scratch.

639
00:23:29,720 --> 00:23:33,000
You have legacy systems that need to move toward a new architecture.

640
00:23:33,000 --> 00:23:34,840
This agent understands those patterns.

641
00:23:34,840 --> 00:23:36,120
It suggests the best approach

642
00:23:36,120 --> 00:23:38,440
and can even help with the actual refactoring.

643
00:23:38,440 --> 00:23:41,080
The code review agent improves quality at scale.

644
00:23:41,080 --> 00:23:42,920
When humans review every pull request,

645
00:23:42,920 --> 00:23:44,840
the process is slow and inconsistent.

646
00:23:44,840 --> 00:23:46,840
You can add an agent to do the first pass.

647
00:23:46,840 --> 00:23:49,480
It checks for security floors and performance issues.

648
00:23:49,480 --> 00:23:51,080
Humans still do the final review,

649
00:23:51,080 --> 00:23:53,400
but they are looking at code that has already been cleaned up.

650
00:23:53,400 --> 00:23:54,920
The pattern here is clarity.

651
00:23:54,920 --> 00:23:57,640
You are not asking which agent is the most advanced.

652
00:23:57,640 --> 00:24:00,440
You are asking what specific problem you need to solve.

653
00:24:00,440 --> 00:24:02,600
And which agent is built for that task?

654
00:24:02,600 --> 00:24:07,080
The decision framework, Dynamics 365 and M365 agents,

655
00:24:07,080 --> 00:24:09,560
as your agents operate on your infrastructure.

656
00:24:09,560 --> 00:24:13,000
But Dynamics and Microsoft 365 agents operate on your business.

657
00:24:13,000 --> 00:24:16,040
That distinction matters because the decision point is different.

658
00:24:16,040 --> 00:24:18,920
You aren't asking what technical problem you need to solve.

659
00:24:18,920 --> 00:24:21,480
You're asking where intelligence is already sitting inside

660
00:24:21,480 --> 00:24:23,000
the way your people actually work.

661
00:24:23,000 --> 00:24:24,760
Take the sales qualification agent.

662
00:24:24,760 --> 00:24:27,400
It exists because your sales team is drowning in leads.

663
00:24:27,400 --> 00:24:28,760
Your inbound volume is high,

664
00:24:28,760 --> 00:24:31,000
but most of those leads are low probability.

665
00:24:31,000 --> 00:24:33,080
Your reps spend their entire day assessing people

666
00:24:33,080 --> 00:24:34,280
instead of selling to them.

667
00:24:34,280 --> 00:24:37,480
This agent looks at every lead against your specific criteria.

668
00:24:37,480 --> 00:24:39,720
It understands what a qualified opportunity looks like

669
00:24:39,720 --> 00:24:40,680
for your business.

670
00:24:40,680 --> 00:24:43,320
It separates the high probability wins from the noise.

671
00:24:43,320 --> 00:24:45,080
Because the agent handles the filtering,

672
00:24:45,080 --> 00:24:47,560
your reps can finally focus on the opportunities

673
00:24:47,560 --> 00:24:49,160
that are actually worth their time.

674
00:24:49,160 --> 00:24:51,960
The account reconciliation agent solves a different problem.

675
00:24:51,960 --> 00:24:54,840
Finance teams spend hours matching invoices to payments.

676
00:24:54,840 --> 00:24:57,160
Operations teams reconcile orders to shipments.

677
00:24:57,160 --> 00:24:58,280
It is mechanical work.

678
00:24:58,280 --> 00:24:59,560
It is time consuming.

679
00:24:59,560 --> 00:25:02,200
And because humans get tired doing repetitive tasks,

680
00:25:02,200 --> 00:25:03,880
it is incredibly error prone.

681
00:25:03,880 --> 00:25:05,720
The agent understands your matching logic.

682
00:25:05,720 --> 00:25:07,320
It knows what a valid match looks like

683
00:25:07,320 --> 00:25:09,560
and flags the exceptions for a human to review.

684
00:25:09,560 --> 00:25:11,720
This reduces hours of pure manual labor.

685
00:25:11,720 --> 00:25:13,560
Now your finance team audits the results

686
00:25:13,560 --> 00:25:15,640
instead of doing the reconciliation themselves.

687
00:25:15,640 --> 00:25:18,440
The customer intent agent listens across every interaction.

688
00:25:18,440 --> 00:25:19,800
A customer sends an email.

689
00:25:19,800 --> 00:25:22,760
They call support. They fill out a form or post on social media.

690
00:25:22,760 --> 00:25:25,480
Those interactions contain signals about what they actually want

691
00:25:25,480 --> 00:25:27,400
even if they don't say it explicitly.

692
00:25:27,400 --> 00:25:30,120
The intent agent analyzes those signals to surface patterns.

693
00:25:30,120 --> 00:25:33,080
It helps your team understand what the customer is signaling

694
00:25:33,080 --> 00:25:34,680
so you can respond at the right time.

695
00:25:34,680 --> 00:25:36,360
You aren't just answering a question.

696
00:25:36,360 --> 00:25:37,960
You're understanding what is really needed.

697
00:25:37,960 --> 00:25:39,720
The supplier communications agent handles

698
00:25:39,720 --> 00:25:41,400
a very specific friction point.

699
00:25:41,400 --> 00:25:43,800
Supplier interactions are frequent and operational.

700
00:25:43,800 --> 00:25:45,960
They usually follow the same repeatable patterns.

701
00:25:45,960 --> 00:25:47,640
Someone asks to expedite an order.

702
00:25:47,640 --> 00:25:50,200
Someone else wants a delivery status or an inventory check

703
00:25:50,200 --> 00:25:51,560
on a specific SKU.

704
00:25:51,560 --> 00:25:53,400
The agent handles these routine inquiries

705
00:25:53,400 --> 00:25:56,280
and only escalates the genuinely complex issues to a person.

706
00:25:56,280 --> 00:25:57,880
It cuts out the operational overhead

707
00:25:57,880 --> 00:25:59,800
of that constant back and forth.

708
00:25:59,800 --> 00:26:02,040
The field service agent coordinates complexity.

709
00:26:02,040 --> 00:26:03,720
When a technician is dispatched to a job,

710
00:26:03,720 --> 00:26:06,040
the agent pulls all the contacts together.

711
00:26:06,040 --> 00:26:08,360
It looks at the service history on the equipment,

712
00:26:08,360 --> 00:26:09,800
known issues with that model,

713
00:26:09,800 --> 00:26:11,240
and the availability of parts.

714
00:26:11,240 --> 00:26:13,080
It even estimates the time to repair.

715
00:26:13,080 --> 00:26:14,920
The agent coordinates the next appointment

716
00:26:14,920 --> 00:26:16,840
before the current one is even finished.

717
00:26:16,840 --> 00:26:18,520
It ensures the follow-up actually happens.

718
00:26:18,520 --> 00:26:20,360
This reduces the coordination overhead

719
00:26:20,360 --> 00:26:23,400
that usually requires a dispatch team to manage by hand.

720
00:26:23,400 --> 00:26:25,320
Agents in Word, Excel, and PowerPoint

721
00:26:25,320 --> 00:26:27,160
embed intelligence directly into the tools

722
00:26:27,160 --> 00:26:28,360
where knowledge work happens.

723
00:26:28,360 --> 00:26:29,560
If a writer is working in Word,

724
00:26:29,560 --> 00:26:31,720
the agent understands the tone and the audience.

725
00:26:31,720 --> 00:26:34,840
It helps them without pulling them out of their flow.

726
00:26:34,840 --> 00:26:36,760
If an analyst is building a spreadsheet,

727
00:26:36,760 --> 00:26:38,680
the agent understands the data structure.

728
00:26:38,680 --> 00:26:40,920
It helps build formulas, finds errors,

729
00:26:40,920 --> 00:26:42,920
and suggests the right visualizations.

730
00:26:42,920 --> 00:26:44,760
A presenter building slides can use an agent

731
00:26:44,760 --> 00:26:47,400
to maintain narrative flow and story coherence.

732
00:26:47,400 --> 00:26:49,080
None of this requires leaving the app.

733
00:26:49,080 --> 00:26:50,360
None of it breaks your focus.

734
00:26:50,360 --> 00:26:53,240
Co-Pilot Studio is the catch-all for custom agents.

735
00:26:53,240 --> 00:26:55,640
When your business process doesn't fit a pre-built pattern,

736
00:26:55,640 --> 00:26:56,680
you build it here.

737
00:26:56,680 --> 00:26:58,520
If your workflow is specific to your industry

738
00:26:58,520 --> 00:26:59,880
or how your company operates,

739
00:26:59,880 --> 00:27:01,000
you can create what you need

740
00:27:01,000 --> 00:27:03,160
without writing code or hiring developers.

741
00:27:03,160 --> 00:27:06,280
It allows business teams to build for their own specific needs.

742
00:27:06,280 --> 00:27:07,800
The pattern here is fundamentally different

743
00:27:07,800 --> 00:27:08,840
from the Azure side.

744
00:27:08,840 --> 00:27:11,480
Azure agents solve infrastructure problems.

745
00:27:11,480 --> 00:27:13,640
Dynamics and M365 agents

746
00:27:13,640 --> 00:27:17,160
embed intelligence into the systems your teams use every single day.

747
00:27:17,160 --> 00:27:18,760
You aren't bolting on new tools.

748
00:27:18,760 --> 00:27:20,840
You're making your existing tools smarter.

749
00:27:20,840 --> 00:27:23,160
You aren't asking teams to learn a new interface.

750
00:27:23,160 --> 00:27:26,040
You're putting the help exactly where the work is already happening.

751
00:27:26,040 --> 00:27:27,480
That is why the scale is differently.

752
00:27:27,480 --> 00:27:30,840
An IT ops team might have to learn a new platform like Foundry

753
00:27:30,840 --> 00:27:33,000
but a sales rep doesn't have to learn anything new at all.

754
00:27:33,000 --> 00:27:35,480
Their existing CRM just suddenly has intelligence.

755
00:27:35,480 --> 00:27:37,960
A finance person doesn't need software training

756
00:27:37,960 --> 00:27:40,440
because their spreadsheet now helps them work faster.

757
00:27:40,440 --> 00:27:42,360
Your support team doesn't need a new system.

758
00:27:42,360 --> 00:27:44,360
Their existing tools just got better.

759
00:27:44,360 --> 00:27:46,760
The advantage is obvious but the challenge is deeper.

760
00:27:46,760 --> 00:27:48,600
Embedding intelligence into business systems

761
00:27:48,600 --> 00:27:51,240
requires a deep understanding of how those systems work.

762
00:27:51,240 --> 00:27:52,840
You have to understand the workflows,

763
00:27:52,840 --> 00:27:55,400
the data models, and the rules your people live by.

764
00:27:55,400 --> 00:27:56,920
A generic agent cannot do that.

765
00:27:56,920 --> 00:27:59,560
The agent has to be built for your specific business.

766
00:27:59,560 --> 00:28:01,720
This is why governance and orchestration matter.

767
00:28:01,720 --> 00:28:03,800
You aren't just running one or two tools anymore.

768
00:28:03,800 --> 00:28:05,720
You're managing dozens of specialized agents

769
00:28:05,720 --> 00:28:07,160
across your entire company.

770
00:28:07,160 --> 00:28:09,320
Without a framework, that becomes chaos.

771
00:28:09,320 --> 00:28:10,760
The shadow agent problem.

772
00:28:10,760 --> 00:28:12,520
Why governance isn't optional?

773
00:28:12,520 --> 00:28:15,160
Most organizations haven't confronted this reality yet.

774
00:28:15,160 --> 00:28:17,720
You likely have agents running in your tenant right now

775
00:28:17,720 --> 00:28:19,480
that IT doesn't even know about.

776
00:28:19,480 --> 00:28:21,640
This isn't because IT is bad at their job.

777
00:28:21,640 --> 00:28:24,360
It's because building agents has become trivially easy

778
00:28:24,360 --> 00:28:25,960
while governance remains hard.

779
00:28:25,960 --> 00:28:28,600
Teams would rather build first and ask for permission never.

780
00:28:28,600 --> 00:28:30,200
These agents are hiding everywhere.

781
00:28:30,200 --> 00:28:32,120
Someone spins up a co-pilot studio environment

782
00:28:32,120 --> 00:28:33,720
to build a tool for their department.

783
00:28:33,720 --> 00:28:36,600
They don't tell IT because they don't see it as infrastructure.

784
00:28:36,600 --> 00:28:39,560
To them, it's just a feature in a platform they already use.

785
00:28:39,560 --> 00:28:42,440
A power platform team creates a bot for their own automation.

786
00:28:42,440 --> 00:28:45,080
A Teams channel has a bot handling admin tasks.

787
00:28:45,080 --> 00:28:47,080
Someone else is running a custom Python script

788
00:28:47,080 --> 00:28:48,280
that calls an API.

789
00:28:48,280 --> 00:28:50,440
Even your vendors are bringing their own agents bundled

790
00:28:50,440 --> 00:28:51,320
into their software.

791
00:28:51,320 --> 00:28:53,400
By the time IT realizes these exist,

792
00:28:53,400 --> 00:28:55,240
dozens of them are already in production.

793
00:28:55,240 --> 00:28:57,000
They are accessing data and making decisions,

794
00:28:57,000 --> 00:28:59,320
but nobody is officially responsible for them.

795
00:28:59,320 --> 00:29:01,160
The risks start to compound quickly.

796
00:29:01,160 --> 00:29:03,720
You might have an agent reading your entire sharepoint tenant

797
00:29:03,720 --> 00:29:05,640
because someone granted it broad permissions

798
00:29:05,640 --> 00:29:07,480
without thinking about security.

799
00:29:07,480 --> 00:29:09,800
Another agent might be sending emails from a shared mailbox

800
00:29:09,800 --> 00:29:11,800
with no audit trail to show who actually sent them.

801
00:29:11,800 --> 00:29:15,160
You could have agents calling APIs with credentials stored in plain text.

802
00:29:15,160 --> 00:29:17,560
Often, a dozen different agents are doing the same work,

803
00:29:17,560 --> 00:29:20,040
but nobody knows it because they aren't registered anywhere.

804
00:29:20,040 --> 00:29:21,960
The access restricted data duplicate effort

805
00:29:21,960 --> 00:29:23,960
and create massive data quality problems.

806
00:29:23,960 --> 00:29:26,440
This happens because the friction is asymmetrical.

807
00:29:26,440 --> 00:29:27,960
Building an agent is frictionless.

808
00:29:27,960 --> 00:29:29,560
It takes hours instead of weeks.

809
00:29:29,560 --> 00:29:31,080
Governance is the exact opposite.

810
00:29:31,080 --> 00:29:34,680
It requires planning, coordination, and policy definitions.

811
00:29:34,680 --> 00:29:36,840
Most organizations haven't built that process yet,

812
00:29:36,840 --> 00:29:39,160
so teams take the path of least resistance.

813
00:29:39,160 --> 00:29:40,840
They build, they test, and they deploy.

814
00:29:40,840 --> 00:29:43,080
They just skip the governance part entirely.

815
00:29:43,080 --> 00:29:46,040
This is where agent 365 changes the structure of the problem.

816
00:29:46,040 --> 00:29:49,320
It doesn't require you to have a perfect governance process in place before you start.

817
00:29:49,320 --> 00:29:50,680
Instead, it provides visibility.

818
00:29:50,680 --> 00:29:52,680
It discovers the agents that are already running.

819
00:29:52,680 --> 00:29:54,120
Every agent talking to enter,

820
00:29:54,120 --> 00:29:56,600
every agent accessing M365 data,

821
00:29:56,600 --> 00:29:58,840
and every bot in the power platform is surfaced.

822
00:29:58,840 --> 00:30:00,840
It doesn't depend on teams registering themselves.

823
00:30:00,840 --> 00:30:02,360
It finds what already exists.

824
00:30:02,360 --> 00:30:04,040
But discovery is only half the battle.

825
00:30:04,040 --> 00:30:06,600
Once you find an agent, you have to be able to control it.

826
00:30:06,600 --> 00:30:09,800
The Entra agent ID requirement changes the baseline for security.

827
00:30:09,800 --> 00:30:12,440
Every agent gets its own identity in your directory.

828
00:30:12,440 --> 00:30:14,600
When IT finds an agent with too much access,

829
00:30:14,600 --> 00:30:17,160
they don't have to shut it down or beg the creators to fix it.

830
00:30:17,160 --> 00:30:19,160
It can scope that identity themselves.

831
00:30:19,160 --> 00:30:21,960
They can apply access controls directly to the agent.

832
00:30:21,960 --> 00:30:24,520
If it only needs to read one specific SharePoint library,

833
00:30:24,520 --> 00:30:27,080
IT grants access to that library and nothing else.

834
00:30:27,080 --> 00:30:29,880
The agent can't accidentally or maliciously touch anything else.

835
00:30:29,880 --> 00:30:31,080
The scope is locked.

836
00:30:31,080 --> 00:30:32,680
If an agent starts misbehaving,

837
00:30:32,680 --> 00:30:34,440
you can revoke its access instantly.

838
00:30:34,440 --> 00:30:36,760
You don't need to have a meeting with the team that built it.

839
00:30:36,760 --> 00:30:38,440
You just revoke the identity.

840
00:30:38,440 --> 00:30:40,360
The agent stops working right then and there.

841
00:30:40,360 --> 00:30:43,240
Every single action that agent takes is logged under its own identity.

842
00:30:43,240 --> 00:30:45,240
When something goes wrong, your logs won't just show

843
00:30:45,240 --> 00:30:47,320
that a system account accessed the data.

844
00:30:47,320 --> 00:30:50,040
You will see exactly which agent did it and when it happened.

845
00:30:50,040 --> 00:30:51,880
The audit trail leads directly to the source.

846
00:30:51,880 --> 00:30:54,280
This is why governance isn't an optional add-on.

847
00:30:54,280 --> 00:30:57,560
It isn't something you bolt on after you have 50 agents running.

848
00:30:57,560 --> 00:30:59,320
It is the foundation that makes visibility

849
00:30:59,320 --> 00:31:01,080
and control possible in the first place.

850
00:31:01,080 --> 00:31:02,520
You cannot govern what you cannot see.

851
00:31:02,520 --> 00:31:04,440
Agent 365 makes those agents visible.

852
00:31:04,440 --> 00:31:06,200
You cannot have control without identity.

853
00:31:06,200 --> 00:31:09,000
Entra agent ID makes those agents controllable.

854
00:31:09,000 --> 00:31:11,400
Together, these tools solve the shadow agent problem.

855
00:31:11,400 --> 00:31:13,240
The goal isn't to stop teams from building.

856
00:31:13,240 --> 00:31:15,480
The goal is to make sure those agents are manageable,

857
00:31:15,480 --> 00:31:17,400
auditable and secure by design.

858
00:31:17,400 --> 00:31:19,960
If you skip this piece, you are building a fragile system.

859
00:31:19,960 --> 00:31:22,920
Foundry agent service, the production runtime.

860
00:31:22,920 --> 00:31:24,680
Governance is the control plane.

861
00:31:24,680 --> 00:31:27,480
But control only matters if there is something to manage.

862
00:31:27,480 --> 00:31:29,480
The runtime is what gives governance a purpose.

863
00:31:29,480 --> 00:31:32,760
It is the infrastructure where your agents actually live and operate.

864
00:31:32,760 --> 00:31:35,160
Foundry agent service is a managed platform.

865
00:31:35,160 --> 00:31:37,560
In practice, this means the division of labor is very clear.

866
00:31:37,560 --> 00:31:38,360
You write the logic.

867
00:31:38,360 --> 00:31:40,200
You define the goals, you pick the tools,

868
00:31:40,200 --> 00:31:41,880
then you deploy it into Foundry.

869
00:31:41,880 --> 00:31:43,080
Microsoft handles the rest.

870
00:31:43,080 --> 00:31:44,920
They manage the scaling, the isolation,

871
00:31:44,920 --> 00:31:47,080
the networking and the persistent state.

872
00:31:47,080 --> 00:31:48,680
You focus on how the agent behaves

873
00:31:48,680 --> 00:31:51,880
while the platform handles the operational heavy lifting.

874
00:31:51,880 --> 00:31:53,240
Agents run in hosted sessions.

875
00:31:53,240 --> 00:31:55,000
Every session gets its own sandbox.

876
00:31:55,000 --> 00:31:56,600
When a user starts a conversation,

877
00:31:56,600 --> 00:31:59,000
that interaction is placed in an isolated container.

878
00:31:59,400 --> 00:32:02,680
The agent can read and write files within that specific session.

879
00:32:02,680 --> 00:32:04,520
It maintains state across multiple turns.

880
00:32:04,520 --> 00:32:07,000
If the session ends, that local state disappears.

881
00:32:07,000 --> 00:32:08,840
But you can design agents to remember things

882
00:32:08,840 --> 00:32:11,400
across sessions if the use case requires it.

883
00:32:11,400 --> 00:32:14,280
This isolation ensures that if one agent misbehaves,

884
00:32:14,280 --> 00:32:16,440
it won't crash your other executions.

885
00:32:16,440 --> 00:32:19,480
The responses API is the single entry point for everything.

886
00:32:19,480 --> 00:32:21,240
When a system needs to call an agent,

887
00:32:21,240 --> 00:32:22,600
it hits this one endpoint.

888
00:32:22,600 --> 00:32:25,240
The API is agnostic to how you build the agent.

889
00:32:25,240 --> 00:32:26,600
It works with semantic kernel.

890
00:32:26,600 --> 00:32:27,720
It works with autogen.

891
00:32:27,720 --> 00:32:29,240
It works with the agent framework.

892
00:32:29,240 --> 00:32:31,960
Even custom code is fine as long as it follows the protocol.

893
00:32:31,960 --> 00:32:34,440
This is a big deal because it prevents vendor lock-in.

894
00:32:34,440 --> 00:32:36,200
Different teams can use different frameworks,

895
00:32:36,200 --> 00:32:38,440
but they all converge at the same runtime.

896
00:32:38,440 --> 00:32:40,040
Memory and Foundry is a core feature.

897
00:32:40,040 --> 00:32:41,080
It isn't bolted on.

898
00:32:41,080 --> 00:32:42,520
There are three distinct types.

899
00:32:42,520 --> 00:32:45,080
Procedural memory is how the agent learns workflows.

900
00:32:45,080 --> 00:32:47,400
It does a task once, remembers the steps,

901
00:32:47,400 --> 00:32:49,320
and becomes faster the second time.

902
00:32:49,320 --> 00:32:51,480
User memory stays active across sessions.

903
00:32:51,480 --> 00:32:54,760
The agent learns your preferences and constraints over time.

904
00:32:54,760 --> 00:32:56,600
Session memory is strictly contextual.

905
00:32:56,600 --> 00:32:58,200
It lives for the current conversation

906
00:32:58,200 --> 00:33:00,360
and vanishes the moment you close the window.

907
00:33:00,360 --> 00:33:03,560
This distinction changes how agents actually feel to the user.

908
00:33:03,560 --> 00:33:06,360
A sales agent with user memory knows your industry

909
00:33:06,360 --> 00:33:07,880
and your typical deal size.

910
00:33:07,880 --> 00:33:10,680
It uses that history to calibrate every suggestion it makes.

911
00:33:10,680 --> 00:33:12,920
But session memory keeps the current thread clean.

912
00:33:12,920 --> 00:33:14,040
When you start a new topic,

913
00:33:14,040 --> 00:33:16,520
you aren't fighting irrelevant history from three days ago.

914
00:33:16,520 --> 00:33:18,520
It prevents the context from getting polluted.

915
00:33:18,520 --> 00:33:20,920
Multi-agent orchestration happens naturally here.

916
00:33:20,920 --> 00:33:23,960
One agent can call another agent just like it would call a tool.

917
00:33:23,960 --> 00:33:26,520
You might have a coordinator agent talking to the user.

918
00:33:26,520 --> 00:33:29,560
When a complex request comes in, the coordinator breaks it down.

919
00:33:29,560 --> 00:33:31,160
It calls a retrieval agent for data.

920
00:33:31,160 --> 00:33:33,720
It calls an analysis agent to process that data.

921
00:33:33,720 --> 00:33:36,760
It calls a communication agent to draft the final text.

922
00:33:36,760 --> 00:33:38,840
The runtime manages this entire flow.

923
00:33:38,840 --> 00:33:41,480
Each agent knows what to call and what to expect back.

924
00:33:41,480 --> 00:33:44,760
If a step fails, the coordinator can retry or escalate the issue.

925
00:33:44,760 --> 00:33:47,400
Toolboxes are how agents find new skills at runtime.

926
00:33:47,400 --> 00:33:50,280
You don't have to hard-code every tool into the agent itself.

927
00:33:50,280 --> 00:33:51,960
Instead, you create a managed library.

928
00:33:51,960 --> 00:33:54,280
When an agent hits a problem, it checks the toolbox

929
00:33:54,280 --> 00:33:56,120
and picks the right tool for the job.

930
00:33:56,120 --> 00:33:57,800
When you add a new tool to the library,

931
00:33:57,800 --> 00:33:59,880
every agent gets access to it immediately.

932
00:33:59,880 --> 00:34:02,040
There is no redeployment and no configuration change.

933
00:34:02,040 --> 00:34:03,880
It is just instant availability.

934
00:34:03,880 --> 00:34:07,720
MCP support allows agents to talk to model context protocol servers.

935
00:34:07,720 --> 00:34:10,680
These are external data sources that speak a standard language.

936
00:34:10,680 --> 00:34:14,280
You don't need to build custom connectors for every legacy system in your stack.

937
00:34:14,280 --> 00:34:16,200
The agent just needs to understand MCP.

938
00:34:16,200 --> 00:34:18,840
If a tool exposes that protocol, the agent can use it.

939
00:34:18,840 --> 00:34:21,240
Private networking is the final piece of the puzzle.

940
00:34:21,240 --> 00:34:24,840
For regulated industries, public endpoints are a non-starter.

941
00:34:24,840 --> 00:34:27,160
Foundry supports end-to-end private networking.

942
00:34:27,160 --> 00:34:29,400
Your agents run entirely inside your vnet.

943
00:34:29,400 --> 00:34:31,880
There are no public endpoints and no outside exposure.

944
00:34:31,880 --> 00:34:33,640
If you work in banking or healthcare,

945
00:34:33,640 --> 00:34:36,440
this is the feature that makes agent infrastructure possible.

946
00:34:36,440 --> 00:34:37,240
The outcome is simple.

947
00:34:37,240 --> 00:34:38,440
You aren't managing clusters.

948
00:34:38,440 --> 00:34:40,120
You aren't writing scaling logic.

949
00:34:40,120 --> 00:34:43,160
You aren't designing database schemas for conversation state.

950
00:34:43,160 --> 00:34:46,200
You define the behavior and let the platform handle the execution.

951
00:34:46,200 --> 00:34:48,680
Multi-agent orchestration, the real complexity.

952
00:34:48,680 --> 00:34:50,600
The runtime gives you a place to run your code,

953
00:34:50,600 --> 00:34:52,040
but running one agent is easy.

954
00:34:52,040 --> 00:34:53,880
Running dozens of agents that have to cooperate

955
00:34:53,880 --> 00:34:55,160
is where most projects fail.

956
00:34:55,160 --> 00:34:56,760
That coordination is the real challenge.

957
00:34:56,760 --> 00:34:59,240
It isn't that the agents are broken, they work fine.

958
00:34:59,240 --> 00:35:02,200
The problem is that orchestration is structurally difficult.

959
00:35:02,200 --> 00:35:04,840
Most organizations haven't built a model to handle it.

960
00:35:04,840 --> 00:35:07,800
The connected agent's pattern is the simplest place to start.

961
00:35:07,800 --> 00:35:10,280
This is where one agent calls another as a tool.

962
00:35:10,280 --> 00:35:12,520
Think of a coordinator agent facing the user.

963
00:35:12,520 --> 00:35:15,480
It receives a request and breaks it into smaller tasks.

964
00:35:15,480 --> 00:35:19,160
It tells a retrieval agent to get information on a specific topic.

965
00:35:19,160 --> 00:35:22,120
It tells an analysis agent to find patterns in that data.

966
00:35:22,120 --> 00:35:25,000
Finally, it tells a communication agent to write the response.

967
00:35:25,000 --> 00:35:26,760
The coordinator then takes those pieces

968
00:35:26,760 --> 00:35:28,600
and stitches them together for the user.

969
00:35:28,600 --> 00:35:30,680
This model works, but it is sequential.

970
00:35:30,680 --> 00:35:33,400
One agent has to finish before the next one can start.

971
00:35:33,400 --> 00:35:34,840
That is fine for basic tasks,

972
00:35:34,840 --> 00:35:37,640
but for complex work, you are wasting time waiting on steps

973
00:35:37,640 --> 00:35:39,080
that could be happening at once.

974
00:35:39,080 --> 00:35:40,920
This is why we use multi-agent workflows.

975
00:35:40,920 --> 00:35:43,480
This is a stateful layer that sits above the runtime.

976
00:35:43,480 --> 00:35:46,760
It coordinates agents over long, multi-step processes.

977
00:35:46,760 --> 00:35:48,120
It isn't just a chain of calls.

978
00:35:48,120 --> 00:35:50,600
It involves branching logic and parallel execution.

979
00:35:50,600 --> 00:35:54,040
It handles errors and keeps state alive over long periods of time.

980
00:35:54,040 --> 00:35:57,240
Durable orchestration handles the most complex scenarios.

981
00:35:57,240 --> 00:35:59,640
Imagine an agent preparing a legal proposal.

982
00:35:59,640 --> 00:36:01,240
It submits the document for approval,

983
00:36:01,240 --> 00:36:01,960
and then it waits.

984
00:36:01,960 --> 00:36:03,640
This might take days or even weeks.

985
00:36:03,640 --> 00:36:05,160
When a human finally approves it,

986
00:36:05,160 --> 00:36:07,960
the workflow resumes with the full context intact.

987
00:36:07,960 --> 00:36:09,880
The agent doesn't have to reread the file.

988
00:36:09,880 --> 00:36:11,480
The state remembers everything.

989
00:36:11,480 --> 00:36:14,040
The execution just continues as if the pause never happened.

990
00:36:14,040 --> 00:36:15,800
Group chat is a completely different model.

991
00:36:15,800 --> 00:36:17,400
Instead of a top-down coordinator,

992
00:36:17,400 --> 00:36:19,320
multiple agents see the same message thread.

993
00:36:19,320 --> 00:36:20,600
They collaborate and they debate.

994
00:36:20,600 --> 00:36:23,400
The control shifts based on the needs of the conversation.

995
00:36:23,400 --> 00:36:25,960
If the group decides to try one approach and it fails,

996
00:36:25,960 --> 00:36:27,320
they can pivot to another.

997
00:36:27,320 --> 00:36:29,160
They are working the problem out together,

998
00:36:29,160 --> 00:36:31,000
rather than following a fixed script.

999
00:36:31,000 --> 00:36:33,800
The choice between sequential and concurrent is about efficiency.

1000
00:36:33,800 --> 00:36:36,440
You use sequential when step B depends on step A.

1001
00:36:36,440 --> 00:36:38,440
You can't analyze data you haven't found yet.

1002
00:36:38,440 --> 00:36:40,760
But you use concurrent when tasks are independent.

1003
00:36:40,760 --> 00:36:43,640
You can pull data from three different databases at the same time

1004
00:36:43,640 --> 00:36:45,080
and then merge the results.

1005
00:36:45,080 --> 00:36:47,480
A good orchestration layer lets you mix both patterns

1006
00:36:47,480 --> 00:36:48,760
in the same workflow.

1007
00:36:48,760 --> 00:36:50,040
Then you have the handoff pattern.

1008
00:36:50,040 --> 00:36:52,680
When a task moves from one agent to another,

1009
00:36:52,680 --> 00:36:54,120
the transition has to be clean.

1010
00:36:54,120 --> 00:36:55,960
It isn't enough to just pass the data.

1011
00:36:55,960 --> 00:36:57,320
You have to transfer the context.

1012
00:36:57,320 --> 00:37:00,040
The second agent needs to know what the first agent already tried.

1013
00:37:00,040 --> 00:37:01,320
It needs to know what failed,

1014
00:37:01,320 --> 00:37:03,160
so it doesn't repeat the same mistakes.

1015
00:37:03,160 --> 00:37:05,480
It picks up exactly where the last agent left off.

1016
00:37:05,480 --> 00:37:08,600
The real complexity starts when you combine these ideas.

1017
00:37:08,600 --> 00:37:10,760
You might have a workflow that is mostly sequential

1018
00:37:10,760 --> 00:37:12,680
but has parallel steps in the middle.

1019
00:37:12,680 --> 00:37:14,680
You might have agents collaborating in a chat

1020
00:37:14,680 --> 00:37:17,480
while a human approval step pauses the whole thing.

1021
00:37:17,480 --> 00:37:20,360
Context has to flow through all these different patterns without breaking.

1022
00:37:20,360 --> 00:37:23,800
Most deployments fail here because the orchestration wasn't planned.

1023
00:37:23,800 --> 00:37:26,600
Someone builds three great agents that work in isolation

1024
00:37:26,600 --> 00:37:28,760
but getting them to work together with proper,

1025
00:37:28,760 --> 00:37:31,320
error handling and audit trails is the actual work.

1026
00:37:31,320 --> 00:37:32,920
A single agent is a useful tool

1027
00:37:32,920 --> 00:37:35,960
but multiple agents working in a coordinated system are transformative.

1028
00:37:35,960 --> 00:37:38,760
They can solve problems that no single model can touch.

1029
00:37:38,760 --> 00:37:40,440
Understanding these orchestration patterns

1030
00:37:40,440 --> 00:37:42,600
is the difference between a toy and a production system.

1031
00:37:42,600 --> 00:37:46,440
The identity shift, agents as principles, not features.

1032
00:37:46,440 --> 00:37:49,800
The structural foundation of everything we've talked about comes down to one thing.

1033
00:37:49,800 --> 00:37:50,840
Identity.

1034
00:37:50,840 --> 00:37:54,920
And how Microsoft fundamentally changed what an agent identity actually means.

1035
00:37:54,920 --> 00:37:56,920
In most organizations building agents today,

1036
00:37:56,920 --> 00:37:57,960
there's an old pattern.

1037
00:37:57,960 --> 00:38:00,280
The agent runs under a shared service account.

1038
00:38:00,280 --> 00:38:01,960
Maybe it's a generic automation account

1039
00:38:01,960 --> 00:38:03,720
or maybe it's a human user account

1040
00:38:03,720 --> 00:38:06,680
that was mapped to the agent because someone needed a quick solution.

1041
00:38:06,680 --> 00:38:09,560
The account has broad permissions because nobody wanted to be granular.

1042
00:38:09,560 --> 00:38:13,240
When the agent acts, the audit log shows that generic account doing the work.

1043
00:38:13,240 --> 00:38:14,760
You can't see which agent did what.

1044
00:38:14,760 --> 00:38:16,200
You can't revoke just that agent.

1045
00:38:16,200 --> 00:38:18,280
You can't apply policies to just that agent.

1046
00:38:18,280 --> 00:38:19,800
It's invisible in the identity system.

1047
00:38:19,800 --> 00:38:21,640
The new model is fundamentally different.

1048
00:38:21,640 --> 00:38:23,880
Every agent gets its own entry agent ID.

1049
00:38:23,880 --> 00:38:25,880
Not a service account, not a user account.

1050
00:38:25,880 --> 00:38:27,400
It's own principle in your directory.

1051
00:38:27,400 --> 00:38:29,400
It shows up in an entry like any other actor.

1052
00:38:29,400 --> 00:38:30,600
It has its own credentials.

1053
00:38:30,600 --> 00:38:31,480
It's own audit trail.

1054
00:38:31,480 --> 00:38:33,240
It's own life cycle.

1055
00:38:33,240 --> 00:38:34,760
That sounds like a technical detail.

1056
00:38:34,760 --> 00:38:37,560
It's actually the foundation that makes governance possible.

1057
00:38:37,560 --> 00:38:38,440
Start with credentials.

1058
00:38:38,440 --> 00:38:40,440
In the old model, agents either had credentials

1059
00:38:40,440 --> 00:38:42,120
hard coded in configuration files

1060
00:38:42,120 --> 00:38:45,160
or they inherited them from whatever account was running the process.

1061
00:38:45,160 --> 00:38:46,600
Both approaches are vulnerable.

1062
00:38:46,600 --> 00:38:49,320
Hard coded credentials and files get checked into repositories

1063
00:38:49,320 --> 00:38:51,720
where they get exposed and shared across multiple agents.

1064
00:38:51,720 --> 00:38:53,160
If one agent is compromised,

1065
00:38:53,160 --> 00:38:56,120
the credential is compromised for every single agent using it.

1066
00:38:56,120 --> 00:38:57,160
With EntraAgentID,

1067
00:38:57,160 --> 00:38:59,240
credential management changes completely.

1068
00:38:59,240 --> 00:39:00,840
The agent doesn't store credentials.

1069
00:39:00,840 --> 00:39:02,040
It doesn't even know them.

1070
00:39:02,040 --> 00:39:03,640
It requests a token from Azure.

1071
00:39:03,640 --> 00:39:05,720
Azure validates the agent's identity.

1072
00:39:05,720 --> 00:39:08,120
If the identity is legitimate and hasn't been revoked,

1073
00:39:08,120 --> 00:39:09,240
the agent gets a token.

1074
00:39:09,240 --> 00:39:10,440
That token is short-lived.

1075
00:39:10,440 --> 00:39:11,240
It expires.

1076
00:39:11,240 --> 00:39:12,760
It's tied to that specific agent.

1077
00:39:12,760 --> 00:39:14,120
If the agent is compromised,

1078
00:39:14,120 --> 00:39:15,240
the token is compromised,

1079
00:39:15,240 --> 00:39:17,720
but only for that agent, other agents are unaffected.

1080
00:39:17,720 --> 00:39:20,440
More importantly, credentials stay in Azure Key Vault.

1081
00:39:20,440 --> 00:39:21,960
They aren't in files or code.

1082
00:39:21,960 --> 00:39:24,680
They're in a secured vault that's separate from the agent itself.

1083
00:39:24,680 --> 00:39:26,760
If the agent requests what it needs from Key Vault

1084
00:39:26,760 --> 00:39:28,520
which then validates the agent's identity

1085
00:39:28,520 --> 00:39:29,800
and returns the secret.

1086
00:39:29,800 --> 00:39:31,080
If you need to rotate a credential,

1087
00:39:31,080 --> 00:39:32,120
you do it in Key Vault.

1088
00:39:32,120 --> 00:39:35,240
Every agent that uses that secret automatically gets the new version.

1089
00:39:35,240 --> 00:39:37,240
No redeployment, no configuration changes.

1090
00:39:37,240 --> 00:39:39,240
The audit trail changes too.

1091
00:39:39,240 --> 00:39:41,960
Every action an agent takes is logged with its identity.

1092
00:39:41,960 --> 00:39:44,920
Not a generic account, not a human user, the specific agent.

1093
00:39:44,920 --> 00:39:49,080
When an audit log shows agent X access this data at this time,

1094
00:39:49,080 --> 00:39:50,680
you know exactly which agent did it.

1095
00:39:50,680 --> 00:39:52,120
You can trace back to who built it,

1096
00:39:52,120 --> 00:39:53,800
who deployed it, and who owns it.

1097
00:39:53,800 --> 00:39:56,040
For regulated industries, this is transformative.

1098
00:39:56,040 --> 00:39:57,240
You can prove compliance.

1099
00:39:57,240 --> 00:39:59,160
You can show that access was controlled.

1100
00:39:59,160 --> 00:40:01,080
You can demonstrate that unintended access

1101
00:40:01,080 --> 00:40:03,560
was impossible because the agent's permissions were scope

1102
00:40:03,560 --> 00:40:04,360
to prevent it.

1103
00:40:04,360 --> 00:40:05,400
Revocation is immediate.

1104
00:40:05,400 --> 00:40:08,280
If an agent is compromised or misbehaving,

1105
00:40:08,280 --> 00:40:09,720
you revoke its identity.

1106
00:40:09,720 --> 00:40:13,320
Not let's shut down the process or let's escalate this to the team that built it.

1107
00:40:13,320 --> 00:40:15,160
You revoke the identity in Entra.

1108
00:40:15,160 --> 00:40:17,000
The agent stops operating immediately.

1109
00:40:17,000 --> 00:40:20,440
Any request it makes is rejected because its identity is no longer valid.

1110
00:40:20,440 --> 00:40:23,000
Everything downstream knows this agent is revoked.

1111
00:40:23,000 --> 00:40:27,160
Conditional access policies work for agents the same way they work for humans.

1112
00:40:27,160 --> 00:40:31,160
If you have a policy that says access to sensitive data requires a managed device,

1113
00:40:31,160 --> 00:40:33,000
that policy applies to agents too.

1114
00:40:33,000 --> 00:40:36,040
If you've set risk-based access controls, agents are subject to them.

1115
00:40:36,040 --> 00:40:37,880
Access reviews include agents.

1116
00:40:37,880 --> 00:40:42,120
If someone leaves the company, you don't have to hunt down every agent they built and disable it.

1117
00:40:42,120 --> 00:40:45,800
If their identity is in a group that's being reviewed, agents they created show up,

1118
00:40:45,800 --> 00:40:48,200
they get audited, they get disabled if necessary.

1119
00:40:48,200 --> 00:40:49,320
The shift is structural.

1120
00:40:49,320 --> 00:40:52,200
Agents stop being invisible infrastructure

1121
00:40:52,200 --> 00:40:54,680
and become managed principles in your identity system.

1122
00:40:54,680 --> 00:40:55,560
That's why it matters.

1123
00:40:55,560 --> 00:40:57,640
You can't have governance without identity.

1124
00:40:57,640 --> 00:41:00,120
With Entra agent ID, governance becomes possible.

1125
00:41:00,120 --> 00:41:03,640
Security agents, reducing analyst fatigue.

1126
00:41:03,640 --> 00:41:05,960
Now that we understand the structural foundation,

1127
00:41:05,960 --> 00:41:07,480
how agents get identity,

1128
00:41:07,480 --> 00:41:09,880
how they're governed and how they orchestrate,

1129
00:41:09,880 --> 00:41:12,280
let's look at what this actually enables in practice.

1130
00:41:12,280 --> 00:41:17,400
Security is where agents start delivering immediate value because the problem they solve is acute and measurable.

1131
00:41:17,400 --> 00:41:19,320
The problem is straightforward.

1132
00:41:19,320 --> 00:41:22,360
Alert volume in most organizations is overwhelming.

1133
00:41:22,360 --> 00:41:24,760
Your team generates thousands of alerts per day.

1134
00:41:24,760 --> 00:41:26,920
Your email system flags suspicious messages,

1135
00:41:26,920 --> 00:41:29,640
your endpoint detection tools, flag anomalous behavior,

1136
00:41:29,640 --> 00:41:31,560
your vulnerability scanner's find issues.

1137
00:41:31,560 --> 00:41:32,920
The volume is relentless.

1138
00:41:32,920 --> 00:41:35,880
And security analysts who are expensive and in short supply

1139
00:41:35,880 --> 00:41:39,160
spend their time waiting through alerts that are 80% noise.

1140
00:41:39,160 --> 00:41:41,480
They're looking for the 20% that actually matter.

1141
00:41:41,480 --> 00:41:43,320
That's not security work, that's filtering.

1142
00:41:43,320 --> 00:41:47,000
The fishing triage agent handles that filtering at the email layer.

1143
00:41:47,000 --> 00:41:49,880
An organization gets hundreds of suspicious messages every week

1144
00:41:49,880 --> 00:41:52,520
and while some are legitimate, many are phishing attempts.

1145
00:41:52,520 --> 00:41:55,880
The agent analyzes incoming messages flagged as potentially malicious

1146
00:41:55,880 --> 00:41:58,360
by looking at sender reputation and message structure.

1147
00:41:58,360 --> 00:42:00,120
It checks for common fishing indicators

1148
00:42:00,120 --> 00:42:02,360
and separates likely phishing from legitimate mail.

1149
00:42:02,360 --> 00:42:03,880
Not perfectly, nothing is perfect,

1150
00:42:03,880 --> 00:42:06,600
but accurately enough that you're reducing the email analysts'

1151
00:42:06,600 --> 00:42:08,360
review workload significantly.

1152
00:42:08,360 --> 00:42:12,040
Instead of reviewing 50 messages to find five that actually need escalation,

1153
00:42:12,040 --> 00:42:16,200
the analyst reviews 20 because the agent already filtered out the obvious cases.

1154
00:42:16,200 --> 00:42:18,360
Alert triage agents work the same principle

1155
00:42:18,360 --> 00:42:20,200
but across the security infrastructure.

1156
00:42:20,200 --> 00:42:21,960
Your tools are generating alerts.

1157
00:42:21,960 --> 00:42:24,760
Some indicate real threats, some are false positives,

1158
00:42:24,760 --> 00:42:27,160
some are legitimate findings but low priority.

1159
00:42:27,160 --> 00:42:30,360
The alert triage agent understands the relationships between alerts.

1160
00:42:30,360 --> 00:42:32,120
It knows which ones cluster together

1161
00:42:32,120 --> 00:42:33,800
and suggests a larger incident.

1162
00:42:33,800 --> 00:42:36,200
It prioritizes based on risk context.

1163
00:42:36,200 --> 00:42:38,920
It surfaces what actually needs immediate attention.

1164
00:42:38,920 --> 00:42:42,040
If you had a thousand alerts yesterday and 500 required investigation

1165
00:42:42,040 --> 00:42:43,560
but only 50 actually mattered,

1166
00:42:43,560 --> 00:42:47,960
the agent helps security teams see the 50 instead of drowning in the 500.

1167
00:42:47,960 --> 00:42:50,360
The conditional access agent operates differently.

1168
00:42:50,360 --> 00:42:54,680
It's not triaging, it's helping security teams manage identity policy at scale.

1169
00:42:54,680 --> 00:42:57,240
Most organizations have conditional access policies.

1170
00:42:57,240 --> 00:43:00,440
Rules that say things like require multi-factor authentication

1171
00:43:00,440 --> 00:43:04,760
for sensitive data access or block access from unusual locations

1172
00:43:04,760 --> 00:43:06,600
but conditional access is complex.

1173
00:43:06,600 --> 00:43:08,760
A change to one policy can have cascading effects

1174
00:43:08,760 --> 00:43:10,600
across other policies and systems.

1175
00:43:10,600 --> 00:43:12,840
The agent helps teams understand policy impact

1176
00:43:12,840 --> 00:43:14,280
before they deploy changes.

1177
00:43:14,280 --> 00:43:16,040
What happens if we modify this rule?

1178
00:43:16,040 --> 00:43:17,000
Who gets affected?

1179
00:43:17,000 --> 00:43:17,880
What breaks?

1180
00:43:17,880 --> 00:43:19,560
Where's the risk?

1181
00:43:19,560 --> 00:43:22,840
The agent helps teams operate their identity policies more safely.

1182
00:43:22,840 --> 00:43:26,440
Vulnerability, remediation agents address the backlog problem.

1183
00:43:26,440 --> 00:43:29,880
Every organization has more vulnerabilities than resources to fix them.

1184
00:43:29,880 --> 00:43:31,640
The agent doesn't just find vulnerabilities.

1185
00:43:31,640 --> 00:43:32,840
That's what scanners do.

1186
00:43:32,840 --> 00:43:36,920
It prioritizes by risk, it connects vulnerability data with threat intelligence.

1187
00:43:36,920 --> 00:43:41,240
It understands which vulnerabilities are actually being exploited in the wild right now

1188
00:43:41,240 --> 00:43:43,000
versus theoretical vulnerabilities.

1189
00:43:43,000 --> 00:43:44,280
It helps plan remediation.

1190
00:43:44,280 --> 00:43:45,240
What should we fix first?

1191
00:43:45,240 --> 00:43:46,280
What can we defer?

1192
00:43:46,280 --> 00:43:48,600
What do we need to monitor while we're waiting to patch?

1193
00:43:48,600 --> 00:43:52,200
The agent reduces the time analysts spend prioritizing and planning.

1194
00:43:52,200 --> 00:43:55,880
The threat intelligence agent connects internal signals with external context.

1195
00:43:55,880 --> 00:43:58,520
Your organization has internal logs and observations

1196
00:43:58,520 --> 00:44:01,000
but threat intelligence, information about campaigns,

1197
00:44:01,000 --> 00:44:03,960
threat actors and emerging techniques, lives outside your network.

1198
00:44:03,960 --> 00:44:05,320
The agent pulls those together.

1199
00:44:05,320 --> 00:44:08,680
It helps analysts understand what they're seeing in that external context.

1200
00:44:08,680 --> 00:44:10,040
This alert you're investigating.

1201
00:44:10,040 --> 00:44:11,800
Is it related to a known campaign?

1202
00:44:11,800 --> 00:44:13,880
Are other organizations seeing similar activity?

1203
00:44:13,880 --> 00:44:15,240
What's the likely motivation?

1204
00:44:15,240 --> 00:44:16,280
How should we respond?

1205
00:44:16,280 --> 00:44:17,800
M-Dash is the testing layer.

1206
00:44:17,800 --> 00:44:19,640
Security agents make decisions that matter.

1207
00:44:19,640 --> 00:44:22,520
You need to validate those decisions before they go into production.

1208
00:44:22,520 --> 00:44:25,400
M-Dash is a security harness that tests agent workflows.

1209
00:44:25,400 --> 00:44:26,760
It runs scenarios.

1210
00:44:26,760 --> 00:44:29,320
It validates that agents are making the right calls.

1211
00:44:29,320 --> 00:44:30,760
It ensures consistency.

1212
00:44:30,760 --> 00:44:33,320
The pattern across all of these is identical.

1213
00:44:33,320 --> 00:44:34,760
Analysts are expensive.

1214
00:44:34,760 --> 00:44:36,120
Agents are cheap.

1215
00:44:36,120 --> 00:44:39,000
If an agent can handle routine assessment and prioritization

1216
00:44:39,000 --> 00:44:41,640
accurately enough that analysts focus their time on

1217
00:44:41,640 --> 00:44:44,040
investigation and response instead of filtering

1218
00:44:44,040 --> 00:44:45,560
that changes the economics.

1219
00:44:45,560 --> 00:44:47,400
Not by replacing analysts.

1220
00:44:47,400 --> 00:44:49,960
By letting them do the actual security work.

1221
00:44:49,960 --> 00:44:53,640
Business process agents embedding intelligence where work happens.

1222
00:44:53,640 --> 00:44:56,440
The security example works because the boundaries are clear.

1223
00:44:56,440 --> 00:44:59,880
Security agents live in a specific domain with specific tools

1224
00:44:59,880 --> 00:45:01,160
and specific workflows.

1225
00:45:01,160 --> 00:45:03,880
But most work doesn't happen in a vacuum.

1226
00:45:03,880 --> 00:45:07,400
It happens inside the business systems your teams use every single day.

1227
00:45:07,400 --> 00:45:08,760
That's where the real scale is.

1228
00:45:08,760 --> 00:45:12,760
The shift from security agents to business process agents is fundamental.

1229
00:45:12,760 --> 00:45:15,160
Security agents reduce the noise for an analyst.

1230
00:45:15,160 --> 00:45:17,960
Business process agents actually become part of how the work gets done.

1231
00:45:17,960 --> 00:45:20,280
They aren't separate tools you open when you need a hand.

1232
00:45:20,280 --> 00:45:23,080
They live inside the systems where the work is already happening.

1233
00:45:23,080 --> 00:45:24,200
Take sales qualification.

1234
00:45:25,240 --> 00:45:27,640
Most sales teams are drowning in inbound leads.

1235
00:45:27,640 --> 00:45:31,720
A campaign runs, the leads pour in, and then sales development reps

1236
00:45:31,720 --> 00:45:36,360
spend hours trying to figure out which ones are actually worth passing to an account executive.

1237
00:45:36,360 --> 00:45:38,760
Some leads are high probability, some are low probability,

1238
00:45:38,760 --> 00:45:41,320
but at the right company others are just a waste of time.

1239
00:45:41,320 --> 00:45:43,880
The sales qualification agent sits right in your CRM.

1240
00:45:43,880 --> 00:45:46,520
The moment a lead arrives the agent assesses it.

1241
00:45:46,520 --> 00:45:49,080
It understands your specific sales criteria.

1242
00:45:49,080 --> 00:45:52,680
Deal size, industry, company profile, and use case fit.

1243
00:45:52,680 --> 00:45:55,480
It scores leads against what actually matters to your business.

1244
00:45:55,480 --> 00:45:59,880
The result is that sales reps see the high probability leads in their queue first.

1245
00:45:59,880 --> 00:46:02,680
They spend their time on opportunities that deserve their attention

1246
00:46:02,680 --> 00:46:05,640
rather than triaging every single person who filled out a form.

1247
00:46:05,640 --> 00:46:08,280
A counter-conciliation is similar but it lives in finance.

1248
00:46:08,280 --> 00:46:10,600
Finance teams upload transactions every day.

1249
00:46:10,600 --> 00:46:12,840
They have to match them against existing records.

1250
00:46:12,840 --> 00:46:14,200
Invoice to payment.

1251
00:46:14,200 --> 00:46:15,320
PO to receipt.

1252
00:46:15,320 --> 00:46:16,360
Shipment to invoice.

1253
00:46:16,360 --> 00:46:20,280
Historically this meant someone sitting at a spreadsheet matching rows.

1254
00:46:20,280 --> 00:46:21,880
It's mechanical work. It's slow.

1255
00:46:21,880 --> 00:46:23,320
And it's prone to human error.

1256
00:46:23,320 --> 00:46:26,360
The account reconciliation agent understands the matching logic.

1257
00:46:26,360 --> 00:46:27,960
It knows what a valid match looks like.

1258
00:46:27,960 --> 00:46:31,880
It connects the transaction data and flags the exceptions that don't match automatically.

1259
00:46:31,880 --> 00:46:34,200
The finance team still reviews the work they have to,

1260
00:46:34,200 --> 00:46:37,160
but they're reviewing the outliers and validating the results

1261
00:46:37,160 --> 00:46:39,000
instead of doing the manual labor.

1262
00:46:39,000 --> 00:46:41,240
The customer intent agent operates on a broader scale.

1263
00:46:41,240 --> 00:46:42,440
Your customer calls support.

1264
00:46:42,440 --> 00:46:43,240
They send an email.

1265
00:46:43,240 --> 00:46:44,680
They post on social media.

1266
00:46:44,680 --> 00:46:48,360
Every one of those interactions contains a signal about what they actually want.

1267
00:46:48,360 --> 00:46:49,960
But finding that signal is hard.

1268
00:46:49,960 --> 00:46:53,960
The agent analyzes these interactions across every channel and surfaces the patterns.

1269
00:46:53,960 --> 00:46:56,760
It tells your team what the customer is actually signaling

1270
00:46:56,760 --> 00:46:59,000
underneath what they explicitly asked for.

1271
00:46:59,000 --> 00:47:00,760
It's not just about answering the question.

1272
00:47:00,760 --> 00:47:02,520
It's about understanding what's really needed

1273
00:47:02,520 --> 00:47:05,400
so you can respond with better timing and better solutions.

1274
00:47:05,400 --> 00:47:07,240
Supply Communications works the same way.

1275
00:47:07,240 --> 00:47:10,600
If you're a large buyer, you get supplier questions constantly.

1276
00:47:10,600 --> 00:47:13,160
They want to know if you can expedite an order, what the status is,

1277
00:47:13,160 --> 00:47:15,320
or if you have inventory on a specific SKU.

1278
00:47:15,320 --> 00:47:18,040
The supplier communications agent handles the routine stuff.

1279
00:47:18,040 --> 00:47:20,120
It knows your inventory, it knows your lead times,

1280
00:47:20,120 --> 00:47:21,800
it knows your escalation procedures.

1281
00:47:21,800 --> 00:47:23,640
It answers the straightforward questions.

1282
00:47:23,640 --> 00:47:25,640
When an inquiry requires human judgment,

1283
00:47:25,640 --> 00:47:28,840
like a special request or a complex issue, it hands it off.

1284
00:47:28,840 --> 00:47:32,600
Your supplier relations team can then focus on the actual relationships

1285
00:47:32,600 --> 00:47:35,320
rather than answering the same three questions over and over.

1286
00:47:35,320 --> 00:47:36,840
Field service takes that coordination

1287
00:47:36,840 --> 00:47:38,280
and pushes it out to the edge.

1288
00:47:38,280 --> 00:47:40,040
A technician gets sent to a job.

1289
00:47:40,040 --> 00:47:42,920
The Field Service agent pulls together everything that person needs

1290
00:47:42,920 --> 00:47:44,120
before they even arrive.

1291
00:47:44,120 --> 00:47:45,800
Service history on the equipment.

1292
00:47:45,800 --> 00:47:47,560
Known issues with that specific model,

1293
00:47:47,560 --> 00:47:49,960
current parts inventory, estimated time to repair.

1294
00:47:49,960 --> 00:47:51,560
It can even schedule the next appointment

1295
00:47:51,560 --> 00:47:53,240
while the current one is still happening.

1296
00:47:53,240 --> 00:47:54,600
It flags the follow-up is needed

1297
00:47:54,600 --> 00:47:57,160
and coordinates the logistics around the actual fieldwork.

1298
00:47:57,160 --> 00:47:58,520
The advantage here is obvious.

1299
00:47:58,520 --> 00:48:01,240
These agents live in the systems your teams already use.

1300
00:48:01,240 --> 00:48:02,680
A sales rep stays in the CRM,

1301
00:48:02,680 --> 00:48:04,680
a finance person stays in Excel or the ERP.

1302
00:48:04,680 --> 00:48:06,760
A support person stays in the ticket system.

1303
00:48:06,760 --> 00:48:09,080
The agent helps them without pulling them out of their flow,

1304
00:48:09,080 --> 00:48:10,040
but here's the problem.

1305
00:48:10,040 --> 00:48:11,240
This is harder than it sounds.

1306
00:48:11,240 --> 00:48:12,760
These agents can't be generic.

1307
00:48:12,760 --> 00:48:14,200
They require a deep understanding

1308
00:48:14,200 --> 00:48:15,720
of the business logic and data models

1309
00:48:15,720 --> 00:48:17,560
that are specific to your organization.

1310
00:48:17,560 --> 00:48:19,000
What counts as a qualified lead for you

1311
00:48:19,000 --> 00:48:21,320
might be junk for another company?

1312
00:48:21,320 --> 00:48:23,800
What a valid match looks like in your finance department

1313
00:48:23,800 --> 00:48:25,560
depends on your specific processes.

1314
00:48:25,560 --> 00:48:27,240
The agent has to understand your rules,

1315
00:48:27,240 --> 00:48:29,560
your workflows, and your constraints.

1316
00:48:29,560 --> 00:48:31,000
That's where governance and identity

1317
00:48:31,000 --> 00:48:32,520
become the priority again.

1318
00:48:32,520 --> 00:48:33,800
You aren't just running software.

1319
00:48:33,800 --> 00:48:35,320
You're embedding agents into systems

1320
00:48:35,320 --> 00:48:36,360
where they make recommendations

1321
00:48:36,360 --> 00:48:38,360
and decisions that affect real-world operations.

1322
00:48:38,360 --> 00:48:39,480
They have to be auditable.

1323
00:48:39,480 --> 00:48:40,680
They have to be scoped.

1324
00:48:40,680 --> 00:48:41,960
And they need clear ownership.

1325
00:48:41,960 --> 00:48:43,880
This is the moment where agents stop being

1326
00:48:43,880 --> 00:48:44,840
an optional upgrade.

1327
00:48:44,840 --> 00:48:47,240
This is where they become the way work gets done.

1328
00:48:47,240 --> 00:48:51,400
Integration, how agents connect to everything.

1329
00:48:51,400 --> 00:48:53,560
An agent sitting by itself is useless.

1330
00:48:53,560 --> 00:48:54,920
It doesn't matter how smart it is

1331
00:48:54,920 --> 00:48:56,200
or how well it was built.

1332
00:48:56,200 --> 00:48:57,480
If it can't get to the data

1333
00:48:57,480 --> 00:48:59,000
and the systems it needs to do the work,

1334
00:48:59,000 --> 00:49:00,760
it's just a language model having a conversation.

1335
00:49:00,760 --> 00:49:02,520
That's the integration problem.

1336
00:49:02,520 --> 00:49:04,920
Every organization has data scattered everywhere.

1337
00:49:04,920 --> 00:49:07,400
SharePoint, Teams, Outlook, OneDrive,

1338
00:49:07,400 --> 00:49:09,160
Data Warehouses, DataBases.

1339
00:49:09,160 --> 00:49:11,240
Every single one of those has different APIs,

1340
00:49:11,240 --> 00:49:12,840
different security requirements,

1341
00:49:12,840 --> 00:49:14,600
and different ways to access the data.

1342
00:49:14,600 --> 00:49:16,120
An agent needs to connect to all of it

1343
00:49:16,120 --> 00:49:18,360
without being custom-built for every single system.

1344
00:49:18,360 --> 00:49:20,920
Toolbox is solved this by creating a discovery mechanism.

1345
00:49:20,920 --> 00:49:23,800
Think of a toolbox as a managed library of tools,

1346
00:49:23,800 --> 00:49:25,480
integrations, and data sources.

1347
00:49:25,480 --> 00:49:27,240
It's all catalogued and ready to go.

1348
00:49:27,240 --> 00:49:28,680
When an agent has a task,

1349
00:49:28,680 --> 00:49:30,680
it checks the toolbox to see what's available.

1350
00:49:30,680 --> 00:49:32,280
It picks the right tool and uses it.

1351
00:49:32,280 --> 00:49:33,960
When you add a new tool to the toolbox,

1352
00:49:33,960 --> 00:49:36,120
every agent gets access to it immediately.

1353
00:49:36,120 --> 00:49:37,720
You don't have to redeploy the agent.

1354
00:49:37,720 --> 00:49:39,320
You don't have to change the configuration.

1355
00:49:39,320 --> 00:49:41,320
The new capability is just there on day one.

1356
00:49:41,320 --> 00:49:44,840
Microsoft Graph is the main bridge for your Microsoft 365 data.

1357
00:49:44,840 --> 00:49:47,080
Through Graph and Agent can see Teams conversations,

1358
00:49:47,080 --> 00:49:48,840
read and send emails, check calendars,

1359
00:49:48,840 --> 00:49:50,360
and pull documents from SharePoint.

1360
00:49:50,360 --> 00:49:53,240
Graph provides a single, consistent interface for all of it.

1361
00:49:53,240 --> 00:49:55,080
Instead of needing one connector for Teams

1362
00:49:55,080 --> 00:49:57,240
and another for email, the agent just talks to Graph.

1363
00:49:57,240 --> 00:49:58,920
The complexity is handled underneath.

1364
00:49:58,920 --> 00:50:02,040
Azure AI Search is what powers the knowledge-heavy scenarios.

1365
00:50:02,040 --> 00:50:03,800
If an agent needs to find information

1366
00:50:03,800 --> 00:50:05,240
across a massive data set,

1367
00:50:05,240 --> 00:50:06,920
it doesn't try to read everything at once.

1368
00:50:06,920 --> 00:50:08,280
That would overwhelm the model.

1369
00:50:08,280 --> 00:50:10,600
Instead, Azure AI Search handles the indexing

1370
00:50:10,600 --> 00:50:11,560
and the retrieval.

1371
00:50:11,560 --> 00:50:12,760
The agent asks the question,

1372
00:50:12,760 --> 00:50:14,200
"Search finds the relevant parts."

1373
00:50:14,200 --> 00:50:15,880
The agent uses those parts to think.

1374
00:50:15,880 --> 00:50:18,040
The agent doesn't need to know how the search works.

1375
00:50:18,040 --> 00:50:19,400
It just needs the results.

1376
00:50:19,400 --> 00:50:21,400
Fabric data agents work a bit differently.

1377
00:50:21,400 --> 00:50:22,920
They aren't just for finding files.

1378
00:50:22,920 --> 00:50:24,120
They're analytical.

1379
00:50:24,120 --> 00:50:26,120
If an agent needs to query a data warehouse

1380
00:50:26,120 --> 00:50:27,880
or understand a complex data set,

1381
00:50:27,880 --> 00:50:30,120
the fabric agents handle that heavy lifting.

1382
00:50:30,120 --> 00:50:31,720
They understand the data schemers.

1383
00:50:31,720 --> 00:50:32,680
They write the queries.

1384
00:50:32,680 --> 00:50:35,480
They return the results in a way the agent can actually use.

1385
00:50:35,480 --> 00:50:37,560
For everything else, they are MCP servers.

1386
00:50:37,560 --> 00:50:39,800
The model context protocol creates a common language

1387
00:50:39,800 --> 00:50:41,800
for agents to talk to external tools.

1388
00:50:41,800 --> 00:50:43,400
It doesn't matter if you're connecting to Slack

1389
00:50:43,400 --> 00:50:45,160
or a custom internal system.

1390
00:50:45,160 --> 00:50:47,720
If it has an MCP server, the agent can use it.

1391
00:50:47,720 --> 00:50:49,000
This is a big deal because it means

1392
00:50:49,000 --> 00:50:50,600
you can build an integration once

1393
00:50:50,600 --> 00:50:52,600
and let agents on different platforms use it.

1394
00:50:52,600 --> 00:50:54,040
Custom connectors are for the systems

1395
00:50:54,040 --> 00:50:55,880
that don't fit the standard patterns.

1396
00:50:55,880 --> 00:50:57,960
Maybe you have a legacy system from 10 years ago

1397
00:50:57,960 --> 00:50:59,720
that doesn't have an MCP server.

1398
00:50:59,720 --> 00:51:00,920
You build a custom connector

1399
00:51:00,920 --> 00:51:02,520
that translates between that system

1400
00:51:02,520 --> 00:51:03,800
and what the agent understands.

1401
00:51:03,800 --> 00:51:05,880
Now, every agent can use that legacy system.

1402
00:51:05,880 --> 00:51:08,120
You aren't writing custom code for every agent.

1403
00:51:08,120 --> 00:51:10,280
You're building a connector that the agents can discover

1404
00:51:10,280 --> 00:51:11,800
and use on their own.

1405
00:51:11,800 --> 00:51:13,880
The governance layer sits on top of all of this.

1406
00:51:13,880 --> 00:51:15,640
Your data policies don't stop at the agent.

1407
00:51:15,640 --> 00:51:17,240
They flow through the integrations.

1408
00:51:17,240 --> 00:51:18,920
If a document is marked as confidential,

1409
00:51:18,920 --> 00:51:21,480
the agency is that metadata and knows it's restricted.

1410
00:51:21,480 --> 00:51:23,480
If you've set sensitivity labels in SharePoint,

1411
00:51:23,480 --> 00:51:24,760
the agents respect them.

1412
00:51:24,760 --> 00:51:27,320
They won't expose information they aren't supposed to see.

1413
00:51:27,320 --> 00:51:29,960
The security you've already built doesn't get bypassed.

1414
00:51:29,960 --> 00:51:31,640
It becomes part of how the agent functions.

1415
00:51:31,640 --> 00:51:33,400
This is why integration isn't just a feature.

1416
00:51:33,400 --> 00:51:35,320
It's the point where agents stop being assistance

1417
00:51:35,320 --> 00:51:36,600
and start being useful.

1418
00:51:36,600 --> 00:51:39,000
Without integration, an agent can only talk.

1419
00:51:39,000 --> 00:51:40,840
With integration, an agent can act.

1420
00:51:40,840 --> 00:51:42,280
The knowledge problem.

1421
00:51:42,280 --> 00:51:44,120
Grounding agents in reality.

1422
00:51:44,120 --> 00:51:45,640
Imagine what happens when an agent

1423
00:51:45,640 --> 00:51:47,640
doesn't have real information to work with.

1424
00:51:47,640 --> 00:51:49,160
Someone asks a question.

1425
00:51:49,160 --> 00:51:51,800
The agent generates a response that sounds reasonable.

1426
00:51:51,800 --> 00:51:52,600
The words flow.

1427
00:51:52,600 --> 00:51:53,560
The grammar is correct.

1428
00:51:53,560 --> 00:51:54,600
The tone is confident.

1429
00:51:54,600 --> 00:51:55,960
And the answer is completely made up.

1430
00:51:55,960 --> 00:51:57,000
This is hallucination.

1431
00:51:57,000 --> 00:51:58,360
The agent isn't being deceptive.

1432
00:51:58,360 --> 00:51:59,800
It's not intentionally lying.

1433
00:51:59,800 --> 00:52:01,240
It's just doing what language models do

1434
00:52:01,240 --> 00:52:02,840
when they don't have actual data.

1435
00:52:02,840 --> 00:52:04,680
They generate plausible sounding text

1436
00:52:04,680 --> 00:52:06,680
based on patterns they learned during training.

1437
00:52:06,680 --> 00:52:08,120
The result feels authoritative.

1438
00:52:08,120 --> 00:52:09,240
And it's often wrong.

1439
00:52:09,240 --> 00:52:10,920
In a security context, that's dangerous.

1440
00:52:10,920 --> 00:52:13,880
In a business context, it's expensive.

1441
00:52:13,880 --> 00:52:16,040
A sales agent tells a customer you have inventory

1442
00:52:16,040 --> 00:52:17,320
you don't actually have.

1443
00:52:17,320 --> 00:52:18,760
A finance agent makes assumptions

1444
00:52:18,760 --> 00:52:21,320
about reconciliation rules that turn out to be incorrect.

1445
00:52:21,320 --> 00:52:23,320
A support agent confidently explains a feature

1446
00:52:23,320 --> 00:52:24,600
that actually works differently.

1447
00:52:24,600 --> 00:52:25,880
These aren't minor issues.

1448
00:52:25,880 --> 00:52:27,000
They erode trust.

1449
00:52:27,000 --> 00:52:28,200
They create rework.

1450
00:52:28,200 --> 00:52:31,000
They undermine the entire value proposition of using agents.

1451
00:52:31,000 --> 00:52:34,600
The solution is grounding, retrieval, augmented generation.

1452
00:52:34,600 --> 00:52:37,080
Our ag is the pattern that makes this work.

1453
00:52:37,080 --> 00:52:39,640
Instead of letting an agent generate responses from memory,

1454
00:52:39,640 --> 00:52:42,680
the agent first retrieves actual relevant documents or data.

1455
00:52:42,680 --> 00:52:44,760
Then it answers based on what it actually found.

1456
00:52:44,760 --> 00:52:47,720
The response is grounded in reality, not in pattern matching.

1457
00:52:47,720 --> 00:52:50,440
Foundry IQ is the managed implementation of rag

1458
00:52:50,440 --> 00:52:52,040
across your entire organization.

1459
00:52:52,040 --> 00:52:53,640
It's a unified knowledge plane.

1460
00:52:53,640 --> 00:52:55,880
Documents, data from your warehouse.

1461
00:52:55,880 --> 00:52:57,560
Web content, enterprise sources,

1462
00:52:57,560 --> 00:52:59,240
all indexed and available for retrieval.

1463
00:52:59,240 --> 00:53:00,600
When an agent needs information,

1464
00:53:00,600 --> 00:53:03,000
it doesn't ask you to point it to the right place.

1465
00:53:03,000 --> 00:53:04,440
It searches across everything.

1466
00:53:04,440 --> 00:53:05,560
It finds what's relevant.

1467
00:53:05,560 --> 00:53:07,080
It uses that to answer.

1468
00:53:07,080 --> 00:53:08,920
The difference between old rag approaches

1469
00:53:08,920 --> 00:53:11,240
and Foundry IQ is operational.

1470
00:53:11,240 --> 00:53:13,480
Building rag five years ago meant custom pipelines.

1471
00:53:13,480 --> 00:53:15,960
You'd set up document ingestion, you'd configure indexing,

1472
00:53:15,960 --> 00:53:18,440
you'd build retrieval logic, you'd monitor embeddings.

1473
00:53:18,440 --> 00:53:19,960
It was infrastructure.

1474
00:53:19,960 --> 00:53:21,800
Now, it's a platform feature.

1475
00:53:21,800 --> 00:53:23,320
Upload your documents.

1476
00:53:23,320 --> 00:53:25,080
Foundry IQ handles the rest.

1477
00:53:25,080 --> 00:53:27,080
Multiple agents use the same knowledge base,

1478
00:53:27,080 --> 00:53:29,000
updates flow through automatically.

1479
00:53:29,000 --> 00:53:31,400
Semantic search is what makes retrieval work accurately.

1480
00:53:31,400 --> 00:53:33,400
You don't ask an agent to match keywords

1481
00:53:33,400 --> 00:53:34,520
that fails constantly.

1482
00:53:34,520 --> 00:53:36,600
Someone asks, what's our return policy?

1483
00:53:36,600 --> 00:53:38,440
And a keyword search returns documents

1484
00:53:38,440 --> 00:53:41,160
about policy documents because the word policy matched.

1485
00:53:41,160 --> 00:53:43,480
Semantic search understands meaning.

1486
00:53:43,480 --> 00:53:46,200
It knows that the question is asking about refunds and returns.

1487
00:53:46,200 --> 00:53:47,880
It retrieves the right policy document

1488
00:53:47,880 --> 00:53:49,800
even if that document uses different words.

1489
00:53:49,800 --> 00:53:51,320
The agent gets relevant information,

1490
00:53:51,320 --> 00:53:52,600
not just keyword matches.

1491
00:53:52,600 --> 00:53:54,440
Procedural memory layers on top of this.

1492
00:53:54,440 --> 00:53:56,200
The agent retrieves information once.

1493
00:53:56,200 --> 00:53:57,400
It learns from that retrieval.

1494
00:53:57,400 --> 00:53:59,080
The next time it faces a similar question,

1495
00:53:59,080 --> 00:54:00,600
it doesn't need to retrieve again.

1496
00:54:00,600 --> 00:54:02,440
It remembers the approach it used before.

1497
00:54:02,440 --> 00:54:05,240
This reduces both hallucination and token consumption.

1498
00:54:05,240 --> 00:54:07,240
Less retrieval, less uncertainty.

1499
00:54:07,240 --> 00:54:09,800
But retrieval without governance creates a different problem.

1500
00:54:09,800 --> 00:54:11,400
What if an agent retrieves data?

1501
00:54:11,400 --> 00:54:12,520
It shouldn't have access to.

1502
00:54:12,520 --> 00:54:14,440
What if it bypasses your data classification?

1503
00:54:14,440 --> 00:54:15,640
This is where the governance layer

1504
00:54:15,640 --> 00:54:18,120
becomes operational, not just theoretical.

1505
00:54:18,120 --> 00:54:19,960
Sensitivity labels flow through retrieval.

1506
00:54:19,960 --> 00:54:21,960
If a document is marked confidential

1507
00:54:21,960 --> 00:54:24,680
and your agent doesn't have access to confidential material,

1508
00:54:24,680 --> 00:54:26,280
that document won't be retrieved.

1509
00:54:26,280 --> 00:54:28,680
Access controls don't just prevent direct access.

1510
00:54:28,680 --> 00:54:31,880
They prevent incidental exposure through agent retrieval.

1511
00:54:31,880 --> 00:54:33,640
The data governance you've already built

1512
00:54:33,640 --> 00:54:35,000
doesn't get circumvented.

1513
00:54:35,000 --> 00:54:36,440
Grounding is the difference between

1514
00:54:36,440 --> 00:54:38,280
interesting agents and trustworthy agents.

1515
00:54:38,280 --> 00:54:40,040
Interesting is when the agent sounds smart.

1516
00:54:40,040 --> 00:54:41,960
Trustworthy is when the agent is actually right,

1517
00:54:41,960 --> 00:54:44,440
that matters enormously when agents start making decisions

1518
00:54:44,440 --> 00:54:45,640
that affect your business.

1519
00:54:45,640 --> 00:54:48,760
When agents recommend actions that depend on accurate information,

1520
00:54:48,760 --> 00:54:52,040
when ordered trails depend on agents knowing what they're talking about.

1521
00:54:52,040 --> 00:54:54,840
This is why knowledge integration isn't an optional enhancement.

1522
00:54:54,840 --> 00:54:55,880
It's foundational.

1523
00:54:55,880 --> 00:54:58,680
Agents without grounding are worse than no agents at all.

1524
00:54:58,680 --> 00:54:59,880
They're confident errors.

1525
00:54:59,880 --> 00:55:01,880
With grounding agents become reliable.

1526
00:55:01,880 --> 00:55:05,000
Cost and efficiency, the real ROI.

1527
00:55:05,000 --> 00:55:06,920
Organizations always ask the same question

1528
00:55:06,920 --> 00:55:08,200
before committing to agents.

1529
00:55:08,200 --> 00:55:09,880
We can build these, but should we?

1530
00:55:09,880 --> 00:55:11,320
What's actually the financial case?

1531
00:55:11,320 --> 00:55:12,840
The answer starts with efficiency.

1532
00:55:12,840 --> 00:55:15,240
Foundry's orchestration layer cuts token consumption

1533
00:55:15,240 --> 00:55:18,120
by about 50% compared to naive approaches.

1534
00:55:18,120 --> 00:55:20,280
That matters because tokens are your cost driver.

1535
00:55:20,280 --> 00:55:23,000
Every request to the model, every retrieval, every response,

1536
00:55:23,000 --> 00:55:24,440
each one consumes tokens,

1537
00:55:24,440 --> 00:55:26,360
reduce token consumption in half,

1538
00:55:26,360 --> 00:55:28,280
and you've fundamentally changed the economics.

1539
00:55:28,280 --> 00:55:29,160
How does that work?

1540
00:55:29,160 --> 00:55:31,400
It's not magic. It's orchestration discipline.

1541
00:55:31,400 --> 00:55:33,560
When you're building agents without a managed platform,

1542
00:55:33,560 --> 00:55:35,640
you tend to do things inefficiently.

1543
00:55:35,640 --> 00:55:38,600
You retrieve more data than you need because retrieval is expensive.

1544
00:55:38,600 --> 00:55:41,000
You reprocess the same information multiple times

1545
00:55:41,000 --> 00:55:42,680
because you're not managing state well.

1546
00:55:42,680 --> 00:55:46,360
You make redundant model calls because you don't have clear orchestration patterns.

1547
00:55:46,360 --> 00:55:47,880
A managed platform like Foundry,

1548
00:55:47,880 --> 00:55:49,480
bakes inefficiency patterns.

1549
00:55:49,480 --> 00:55:51,480
State management reduces reprocessing,

1550
00:55:51,480 --> 00:55:53,000
caching prevents redundant calls.

1551
00:55:53,000 --> 00:55:55,080
Orchestration ensures you're retrieving what you need,

1552
00:55:55,080 --> 00:55:56,360
not everything you might need.

1553
00:55:56,360 --> 00:55:58,440
The result is fewer tokens per task.

1554
00:55:58,440 --> 00:56:00,600
Same output, half the consumption.

1555
00:56:00,600 --> 00:56:02,440
Cost transparency is equally important.

1556
00:56:02,440 --> 00:56:04,600
You need granular metrics that show cost per agent,

1557
00:56:04,600 --> 00:56:06,040
per task, per tool call.

1558
00:56:06,040 --> 00:56:08,520
Old approaches buried costs across infrastructure.

1559
00:56:08,520 --> 00:56:10,920
You couldn't tell whether agents were actually efficient

1560
00:56:10,920 --> 00:56:13,640
or whether they were just invisible drains on your budget.

1561
00:56:13,640 --> 00:56:16,280
With granular metrics, you see exactly what's happening.

1562
00:56:16,280 --> 00:56:18,200
Agent A costs 50 cents per execution.

1563
00:56:18,200 --> 00:56:20,360
Agent B costs $2 per execution.

1564
00:56:20,360 --> 00:56:21,240
Now you can ask why.

1565
00:56:21,240 --> 00:56:23,320
Did agent B retrieve more data?

1566
00:56:23,320 --> 00:56:24,920
Did it make more tool calls?

1567
00:56:24,920 --> 00:56:26,440
Is there a pattern worth optimizing?

1568
00:56:26,440 --> 00:56:28,040
You can't manage what you can't measure.

1569
00:56:28,040 --> 00:56:30,920
But the real financial impact comes from what agents replace.

1570
00:56:30,920 --> 00:56:32,120
Look at security co-pilot,

1571
00:56:32,120 --> 00:56:33,880
security analysts are expensive.

1572
00:56:33,880 --> 00:56:36,120
A mid-level analyst costs your organization something

1573
00:56:36,120 --> 00:56:40,200
like $120 to $180,000 per year in full cost.

1574
00:56:40,200 --> 00:56:42,200
The fishing triage agent reduces the time

1575
00:56:42,200 --> 00:56:44,360
and analyst spends reviewing suspicious messages

1576
00:56:44,360 --> 00:56:46,200
by 70 to 80%.

1577
00:56:46,200 --> 00:56:48,760
That analyst now spends time on actual investigation

1578
00:56:48,760 --> 00:56:49,640
instead of filtering.

1579
00:56:49,640 --> 00:56:51,160
That's not replacing the analyst.

1580
00:56:51,160 --> 00:56:53,960
That's deploying them where they actually add value.

1581
00:56:53,960 --> 00:56:55,800
If you have 10 analysts doing triage work

1582
00:56:55,800 --> 00:56:57,560
that could be reduced by 70%,

1583
00:56:57,560 --> 00:56:59,800
you've freed up seven analysts worth of capacity.

1584
00:56:59,800 --> 00:57:02,360
That's roughly $1.2 million in annual cost.

1585
00:57:02,360 --> 00:57:03,320
That's still there.

1586
00:57:03,320 --> 00:57:04,600
The analysts still work.

1587
00:57:04,600 --> 00:57:06,840
But you've deployed them to work that's actually valuable.

1588
00:57:06,840 --> 00:57:08,440
The same math applies to sales.

1589
00:57:08,440 --> 00:57:11,160
Sales development reps spend roughly a quarter of their time

1590
00:57:11,160 --> 00:57:13,480
assessing leads that aren't actually qualified.

1591
00:57:13,480 --> 00:57:15,800
The sales qualification agent handles that assessment.

1592
00:57:15,800 --> 00:57:19,320
75% improvement in rep productivity on lead evaluation

1593
00:57:19,320 --> 00:57:22,280
means more qualified leads getting to account executives faster.

1594
00:57:22,280 --> 00:57:23,480
More deals in the pipeline.

1595
00:57:23,480 --> 00:57:24,520
Same headcount.

1596
00:57:24,520 --> 00:57:25,480
Better output.

1597
00:57:25,480 --> 00:57:27,960
Account reconciliation in finance looks similar.

1598
00:57:27,960 --> 00:57:31,240
Finance staff spend somewhere between 10 and 15 hours per week

1599
00:57:31,240 --> 00:57:32,840
on manual record matching.

1600
00:57:32,840 --> 00:57:35,800
The account reconciliation agent handles that matching.

1601
00:57:35,800 --> 00:57:37,480
You're not eliminating the role,

1602
00:57:37,480 --> 00:57:40,040
but you're eliminating the lowest value part of it.

1603
00:57:40,040 --> 00:57:41,960
The person who was matching records all day

1604
00:57:41,960 --> 00:57:44,280
now orders exceptions and validates results.

1605
00:57:44,280 --> 00:57:45,240
They work smarter.

1606
00:57:45,240 --> 00:57:45,880
Same cost.

1607
00:57:45,880 --> 00:57:47,080
Better use of their time.

1608
00:57:47,080 --> 00:57:49,320
Error reduction has its own cost profile.

1609
00:57:49,320 --> 00:57:50,760
Manual processes fail.

1610
00:57:50,760 --> 00:57:53,240
Someone matches the wrong invoice to the wrong payment.

1611
00:57:53,240 --> 00:57:54,920
Someone assesses a lead incorrectly.

1612
00:57:54,920 --> 00:57:58,040
Someone misses a security alert because they've reviewed too many.

1613
00:57:58,040 --> 00:58:00,120
Those failures create downstream costs.

1614
00:58:00,120 --> 00:58:01,000
Rework.

1615
00:58:01,000 --> 00:58:02,040
Mist opportunities.

1616
00:58:02,040 --> 00:58:03,320
Security incidents.

1617
00:58:03,320 --> 00:58:06,680
An agent that makes fewer errors reduces those downstream costs.

1618
00:58:06,680 --> 00:58:08,520
Not zero errors. Nothing's perfect.

1619
00:58:08,520 --> 00:58:11,000
But better accuracy than humans under time pressure.

1620
00:58:11,000 --> 00:58:12,520
The scaling piece is crucial.

1621
00:58:12,520 --> 00:58:14,520
As volume grows agents scale with it.

1622
00:58:14,520 --> 00:58:15,400
You don't need to hire.

1623
00:58:15,400 --> 00:58:16,520
You don't need more headcount.

1624
00:58:16,520 --> 00:58:17,880
You need more API calls.

1625
00:58:17,880 --> 00:58:19,960
That's a fundamentally different cost structure.

1626
00:58:19,960 --> 00:58:22,760
Your first analyst costs you 150,000.

1627
00:58:22,760 --> 00:58:25,320
Your second analyst costs another 150,000.

1628
00:58:25,320 --> 00:58:27,560
But once you have agents handling baseline work,

1629
00:58:27,560 --> 00:58:29,880
your first and second analyst both cost less

1630
00:58:29,880 --> 00:58:32,920
because they're focusing on work that only humans should do.

1631
00:58:32,920 --> 00:58:34,760
Volume scales without headcount scaling.

1632
00:58:34,760 --> 00:58:37,400
The model shifts from cost per person to cost per task.

1633
00:58:37,400 --> 00:58:39,320
That changes everything financially.

1634
00:58:39,320 --> 00:58:41,960
An analyst can only process so many leads per day.

1635
00:58:41,960 --> 00:58:43,560
An agent processes them instantly.

1636
00:58:43,560 --> 00:58:45,400
That leverage is where the ROI lives.

1637
00:58:45,400 --> 00:58:47,560
This is how you justify agents to CFOs.

1638
00:58:47,560 --> 00:58:49,000
Not AI is cool.

1639
00:58:49,000 --> 00:58:51,160
But this agent saves us this much per month.

1640
00:58:51,160 --> 00:58:52,600
And here's how we measured it.

1641
00:58:52,600 --> 00:58:55,000
Getting started, the practical path.

1642
00:58:55,000 --> 00:58:57,160
You don't need a six month planning cycle to start.

1643
00:58:57,160 --> 00:58:58,600
You need one working agent.

1644
00:58:58,600 --> 00:59:01,800
And the discipline to avoid trying to solve everything at once.

1645
00:59:01,800 --> 00:59:03,400
Stage one is about picking scope.

1646
00:59:03,400 --> 00:59:05,400
Don't say let's use agents everywhere.

1647
00:59:05,400 --> 00:59:08,600
Pick one domain, one process, one clear problem,

1648
00:59:08,600 --> 00:59:10,600
where an agent adds measurable value.

1649
00:59:10,600 --> 00:59:13,800
Maybe it's security triage, or sales qualification,

1650
00:59:13,800 --> 00:59:15,400
or account reconciliation.

1651
00:59:15,400 --> 00:59:17,400
Pick the spot where people are spending time on work

1652
00:59:17,400 --> 00:59:19,800
that's mechanical but still requires judgment.

1653
00:59:19,800 --> 00:59:21,800
That's where agents create immediate impact.

1654
00:59:21,800 --> 00:59:24,600
Copilot Studio is your entry point because it's low code.

1655
00:59:24,600 --> 00:59:25,800
You don't need developers.

1656
00:59:25,800 --> 00:59:27,800
You don't need infrastructure expertise.

1657
00:59:27,800 --> 00:59:29,800
And you definitely don't need months of planning.

1658
00:59:29,800 --> 00:59:32,200
You open it, you describe what you want the agent to do.

1659
00:59:32,200 --> 00:59:33,800
You give it data, you deploy it.

1660
00:59:33,800 --> 00:59:35,400
That's the foundation.

1661
00:59:35,400 --> 00:59:37,400
From the first idea to a working agent,

1662
00:59:37,400 --> 00:59:39,000
you're looking at weeks, not months.

1663
00:59:39,000 --> 00:59:41,400
Ground it specifically in your organization's data.

1664
00:59:41,400 --> 00:59:42,600
This is non-negotiable.

1665
00:59:42,600 --> 00:59:45,400
A generic agent is worthless because it hallucinates.

1666
00:59:45,400 --> 00:59:47,000
Users stop trusting it.

1667
00:59:47,000 --> 00:59:48,600
And it becomes a curiosity instead of a tool.

1668
00:59:48,600 --> 00:59:50,600
But when you upload your company's documents

1669
00:59:50,600 --> 00:59:52,200
and connect your SharePoint libraries,

1670
00:59:52,200 --> 00:59:53,000
something changes.

1671
00:59:53,000 --> 00:59:54,600
The agent understands your business.

1672
00:59:54,600 --> 00:59:56,800
It answers questions about your specific processes

1673
00:59:56,800 --> 00:59:58,200
and your specific rules.

1674
00:59:58,200 --> 00:59:59,400
It's not generic anymore.

1675
00:59:59,400 --> 01:00:00,400
It's useful.

1676
01:00:00,400 --> 01:00:01,400
Don't launch to everyone.

1677
01:00:01,400 --> 01:00:03,000
That's the mistake organizations make.

1678
01:00:03,000 --> 01:00:04,600
They build something in Copilot Studio.

1679
01:00:04,600 --> 01:00:05,400
It works in testing.

1680
01:00:05,400 --> 01:00:06,800
So they enable it for the whole company.

1681
01:00:06,800 --> 01:00:07,800
Now it's in production.

1682
01:00:07,800 --> 01:00:08,800
Now it's breaking.

1683
01:00:08,800 --> 01:00:10,800
Now it's creating problems instead of solving them.

1684
01:00:10,800 --> 01:00:13,000
Instead run it with a small pilot team first.

1685
01:00:13,000 --> 01:00:15,400
10 people, a sales team, a support group.

1686
01:00:15,400 --> 01:00:16,200
Let them use it.

1687
01:00:16,200 --> 01:00:16,800
Get feedback.

1688
01:00:16,800 --> 01:00:17,600
Learn where it breaks.

1689
01:00:17,600 --> 01:00:20,400
Understand what people actually need versus what you build.

1690
01:00:20,400 --> 01:00:21,600
Iterate.

1691
01:00:21,600 --> 01:00:23,200
Once you've done three rounds of feedback

1692
01:00:23,200 --> 01:00:25,000
and improvements, then you start expanding.

1693
01:00:25,000 --> 01:00:26,600
Stage 2 is making it governable.

1694
01:00:26,600 --> 01:00:29,800
Register that agent in agent 365 assign clear ownership.

1695
01:00:29,800 --> 01:00:31,600
Someone responsible for that agent's behavior

1696
01:00:31,600 --> 01:00:33,600
and maintenance set access policies.

1697
01:00:33,600 --> 01:00:34,800
This isn't bureaucracy.

1698
01:00:34,800 --> 01:00:36,600
It's the difference between managing the agent

1699
01:00:36,600 --> 01:00:37,600
and losing control of it.

1700
01:00:37,600 --> 01:00:39,600
Once it's registered, you can revoke access

1701
01:00:39,600 --> 01:00:40,600
if something breaks.

1702
01:00:40,600 --> 01:00:41,800
You can audit what it's doing.

1703
01:00:41,800 --> 01:00:43,800
And you know exactly what data it touches.

1704
01:00:43,800 --> 01:00:44,800
This is the invisible work

1705
01:00:44,800 --> 01:00:46,600
that prevents problems downstream.

1706
01:00:46,600 --> 01:00:48,400
Stage 3 is knowing when to graduate it.

1707
01:00:48,400 --> 01:00:49,600
If the pilot is working.

1708
01:00:49,600 --> 01:00:51,000
If users are actually using it.

1709
01:00:51,000 --> 01:00:53,000
And if it's reducing work in measurable ways.

1710
01:00:53,000 --> 01:00:54,400
That's when you move it to Foundry.

1711
01:00:54,400 --> 01:00:55,800
Not because co-pilot studio is bad,

1712
01:00:55,800 --> 01:00:57,400
it's a perfectly good place to build.

1713
01:00:57,400 --> 01:00:59,000
But Foundry is where you build at scale.

1714
01:00:59,000 --> 01:01:01,000
Co-pilot studio is where you prototype.

1715
01:01:01,000 --> 01:01:02,200
Foundry is where you operate.

1716
01:01:02,200 --> 01:01:03,600
The move gets your orchestration.

1717
01:01:03,600 --> 01:01:04,800
Memory across sessions.

1718
01:01:04,800 --> 01:01:06,600
And multi agent capabilities.

1719
01:01:06,600 --> 01:01:09,200
It's the infrastructure layer that lets a single agent scale

1720
01:01:09,200 --> 01:01:11,200
to thousands of users without degrading.

1721
01:01:11,200 --> 01:01:13,000
Stage 4 is velocity.

1722
01:01:13,000 --> 01:01:14,600
Now you have one agent in production.

1723
01:01:14,600 --> 01:01:15,800
You understand the patterns.

1724
01:01:15,800 --> 01:01:17,000
You've built the governance.

1725
01:01:17,000 --> 01:01:18,200
You've connected the data.

1726
01:01:18,200 --> 01:01:20,000
The second agent is radically faster.

1727
01:01:20,000 --> 01:01:21,800
You reuse the tool integrations.

1728
01:01:21,800 --> 01:01:23,400
You reuse knowledge bases.

1729
01:01:23,400 --> 01:01:25,000
And you reuse governance templates.

1730
01:01:25,000 --> 01:01:27,800
What took weeks the first time takes days the second time.

1731
01:01:27,800 --> 01:01:30,600
By agent number five, you have repeatable patterns.

1732
01:01:30,600 --> 01:01:32,000
You have infrastructure.

1733
01:01:32,000 --> 01:01:33,800
You're not starting from scratch each time.

1734
01:01:33,800 --> 01:01:36,200
The timeline is the part that surprises people.

1735
01:01:36,200 --> 01:01:37,000
Not months.

1736
01:01:37,000 --> 01:01:39,200
Weeks. That's genuinely what managed platforms enable.

1737
01:01:39,200 --> 01:01:40,400
You're not building infrastructure.

1738
01:01:40,400 --> 01:01:42,000
You're not configuring databases.

1739
01:01:42,000 --> 01:01:43,600
And you're not managing containers.

1740
01:01:43,600 --> 01:01:45,600
You're defining logic and letting the platform

1741
01:01:45,600 --> 01:01:47,400
handle the operational complexity.

1742
01:01:47,400 --> 01:01:49,000
None of this requires a perfect plan.

1743
01:01:49,000 --> 01:01:50,000
No organization has that.

1744
01:01:50,000 --> 01:01:51,600
You need a working agent.

1745
01:01:51,600 --> 01:01:53,800
You need the ability to measure if it's working.

1746
01:01:53,800 --> 01:01:56,400
And you need the discipline to improve it before scaling it.

1747
01:01:56,400 --> 01:01:57,800
That's it. Start there.

1748
01:01:57,800 --> 01:01:59,400
Common failures. What not to do?

1749
01:01:59,400 --> 01:02:01,800
Most agent deployments fail in predictable ways.

1750
01:02:01,800 --> 01:02:03,400
Not because the technology doesn't work,

1751
01:02:03,400 --> 01:02:05,400
but because organizations build without thinking

1752
01:02:05,400 --> 01:02:07,400
about what agents actually require.

1753
01:02:07,400 --> 01:02:09,800
Failure one is building without any governance at all.

1754
01:02:09,800 --> 01:02:11,200
You spin up co-pilot studio.

1755
01:02:11,200 --> 01:02:12,200
You build something.

1756
01:02:12,200 --> 01:02:12,800
It works.

1757
01:02:12,800 --> 01:02:14,200
You deploy it to your department.

1758
01:02:14,200 --> 01:02:16,200
Now it's reading from your entire SharePoint tenant

1759
01:02:16,200 --> 01:02:19,000
because you granted broad permissions and never revisited it.

1760
01:02:19,000 --> 01:02:21,000
It's calling APIs with credentials stored

1761
01:02:21,000 --> 01:02:22,400
in configuration files.

1762
01:02:22,400 --> 01:02:24,200
It's accessing data it shouldn't touch.

1763
01:02:24,200 --> 01:02:27,200
And nobody knows it exists except the three people who use it.

1764
01:02:27,200 --> 01:02:29,600
When IT finally discovers it six months later,

1765
01:02:29,600 --> 01:02:31,800
it's running production workloads with zero oversight,

1766
01:02:31,800 --> 01:02:33,200
shutting it down breaks things,

1767
01:02:33,200 --> 01:02:35,600
keeping it running, violates security policy,

1768
01:02:35,600 --> 01:02:37,000
your stuck.

1769
01:02:37,000 --> 01:02:39,400
The agent became a liability instead of an asset

1770
01:02:39,400 --> 01:02:41,800
because governance was never part of the conversation.

1771
01:02:41,800 --> 01:02:44,200
Failure two is different, but equally costly.

1772
01:02:44,200 --> 01:02:47,000
You build an agent without grounding it in actual data.

1773
01:02:47,000 --> 01:02:48,200
The agent sounds intelligent.

1774
01:02:48,200 --> 01:02:50,800
It answers questions fluently and it's making things up.

1775
01:02:50,800 --> 01:02:52,600
When a customer asks about product features,

1776
01:02:52,600 --> 01:02:55,200
it invents specifications that don't exist.

1777
01:02:55,200 --> 01:02:57,200
When finance asks about reconciliation rules,

1778
01:02:57,200 --> 01:02:58,700
it hallucinates policies.

1779
01:02:58,700 --> 01:03:00,000
Users stop trusting it.

1780
01:03:00,000 --> 01:03:01,400
It becomes useless.

1781
01:03:01,400 --> 01:03:02,900
The agent had potential.

1782
01:03:02,900 --> 01:03:04,900
You killed it by skipping the most critical step,

1783
01:03:04,900 --> 01:03:06,600
connecting it to real information,

1784
01:03:06,600 --> 01:03:09,600
treating agents as features instead of systems is failure three.

1785
01:03:09,600 --> 01:03:11,600
You build an agent it works for a few months,

1786
01:03:11,600 --> 01:03:12,600
then something breaks.

1787
01:03:12,600 --> 01:03:13,800
Nobody remembers who built it.

1788
01:03:13,800 --> 01:03:16,100
Nobody knows who owns it and nobody maintains it.

1789
01:03:16,100 --> 01:03:17,900
The data it was trained on is stale.

1790
01:03:17,900 --> 01:03:20,700
The integrations it relied on changed and nobody updated them.

1791
01:03:20,700 --> 01:03:23,200
The agent becomes a ghost running but not useful.

1792
01:03:23,200 --> 01:03:26,100
Too expensive to maintain, but too embedded to shut down.

1793
01:03:26,100 --> 01:03:28,300
This happens because nobody assigned clear ownership,

1794
01:03:28,300 --> 01:03:30,200
nobody planned for life cycle management.

1795
01:03:30,200 --> 01:03:33,200
It was treated as a feature you build once and then forget.

1796
01:03:33,200 --> 01:03:35,100
But agents need care.

1797
01:03:35,100 --> 01:03:36,800
Failure four is over scoping.

1798
01:03:36,800 --> 01:03:39,200
You decide the first agent should solve everything.

1799
01:03:39,200 --> 01:03:41,800
It should handle sales qualification and lead follow-up

1800
01:03:41,800 --> 01:03:44,200
and opportunity scoring and deal analysis.

1801
01:03:44,200 --> 01:03:45,600
It becomes unwieldy.

1802
01:03:45,600 --> 01:03:47,100
It's trying to be good at too many things.

1803
01:03:47,100 --> 01:03:48,700
It's actually good at none of them.

1804
01:03:48,700 --> 01:03:51,200
The agent fails because you gave it an impossible scope.

1805
01:03:51,200 --> 01:03:52,000
Pick one job.

1806
01:03:52,000 --> 01:03:52,700
Do that job well.

1807
01:03:52,700 --> 01:03:55,600
At other jobs later, not measuring impact is failure five.

1808
01:03:55,600 --> 01:03:57,000
You can't justify the investment.

1809
01:03:57,000 --> 01:03:58,700
You don't know if users are actually using it.

1810
01:03:58,700 --> 01:04:01,400
You don't know if it's saving time or just shifting work around.

1811
01:04:01,400 --> 01:04:04,400
You can't answer whether this agent is worth the infrastructure cost.

1812
01:04:04,400 --> 01:04:07,100
You should have defined what success looks like before deployment.

1813
01:04:07,100 --> 01:04:08,800
How many lead assessments per day?

1814
01:04:08,800 --> 01:04:11,100
How much time saved per finance reconciliation?

1815
01:04:11,100 --> 01:04:13,000
Without that baseline, you're running blind.

1816
01:04:13,000 --> 01:04:14,800
Failure six is ignoring orchestration.

1817
01:04:14,800 --> 01:04:16,000
You build three agents.

1818
01:04:16,000 --> 01:04:18,400
They work independently, but they need to work together.

1819
01:04:18,400 --> 01:04:20,100
You don't have orchestration patterns.

1820
01:04:20,100 --> 01:04:22,500
So you're calling them sequentially through manual scripts.

1821
01:04:22,500 --> 01:04:23,900
Each agent does its job.

1822
01:04:23,900 --> 01:04:25,800
Nobody's coordinating them.

1823
01:04:25,800 --> 01:04:27,900
The value multiplies when agents work together.

1824
01:04:27,900 --> 01:04:31,300
If you skip orchestration, you stay stuck at single agent utility.

1825
01:04:31,300 --> 01:04:32,800
You miss the compounding returns.

1826
01:04:32,800 --> 01:04:34,600
Failure seven is not planning for scale.

1827
01:04:34,600 --> 01:04:37,300
An agent works brilliantly when ten people use it.

1828
01:04:37,300 --> 01:04:38,000
You're confident.

1829
01:04:38,000 --> 01:04:39,100
You enable it for everyone.

1830
01:04:39,100 --> 01:04:41,000
Now it's running on ten thousand users.

1831
01:04:41,000 --> 01:04:42,500
Token costs explode.

1832
01:04:42,500 --> 01:04:45,200
Response time degrades.

1833
01:04:45,200 --> 01:04:47,700
The infrastructure that worked for pilots breaks.

1834
01:04:47,700 --> 01:04:49,800
You didn't think about scale from the start.

1835
01:04:49,800 --> 01:04:51,800
You didn't measure token consumption early.

1836
01:04:51,800 --> 01:04:53,200
And you didn't design for concurrency.

1837
01:04:53,200 --> 01:04:55,400
Now you're in crisis mode trying to optimize something

1838
01:04:55,400 --> 01:04:56,900
that was never designed to scale.

1839
01:04:56,900 --> 01:04:59,200
The pattern across all of these is identical.

1840
01:04:59,200 --> 01:05:02,800
Most failures come from treating agents as simple tools

1841
01:05:02,800 --> 01:05:05,100
instead of systems that need governance, maintenance,

1842
01:05:05,100 --> 01:05:06,600
architecture and oversight.

1843
01:05:06,600 --> 01:05:09,700
You wouldn't deploy infrastructure without planning for scale.

1844
01:05:09,700 --> 01:05:11,900
You wouldn't access data without identity and audit.

1845
01:05:11,900 --> 01:05:13,600
You wouldn't skip integration planning.

1846
01:05:13,600 --> 01:05:15,800
But with agents, organizations do all those things

1847
01:05:15,800 --> 01:05:18,700
because agents feel like software features, not infrastructure.

1848
01:05:18,700 --> 01:05:19,500
They're not.

1849
01:05:19,500 --> 01:05:21,200
The sooner you stop treating them that way,

1850
01:05:21,200 --> 01:05:23,300
the sooner your deployments stop failing.

1851
01:05:23,300 --> 01:05:26,400
The organizational shift from assistance to automation.

1852
01:05:26,400 --> 01:05:29,000
Everything we've talked about, the layers, the governance,

1853
01:05:29,000 --> 01:05:31,100
the orchestration, this is all infrastructure,

1854
01:05:31,100 --> 01:05:34,000
but infrastructure only matters because of what it enables.

1855
01:05:34,000 --> 01:05:35,800
And what it enables is fundamentally different

1856
01:05:35,800 --> 01:05:37,600
from how organizations work today.

1857
01:05:37,600 --> 01:05:39,600
Right now, the model is assistance.

1858
01:05:39,600 --> 01:05:40,600
AI helps humans.

1859
01:05:40,600 --> 01:05:41,800
A human sees the results.

1860
01:05:41,800 --> 01:05:45,000
A human makes the decision and a human executes the action.

1861
01:05:45,000 --> 01:05:46,600
The human is still the primary actor.

1862
01:05:46,600 --> 01:05:47,700
The agent is secondary.

1863
01:05:47,700 --> 01:05:50,400
It's just a tool that makes that human faster or more informed.

1864
01:05:50,400 --> 01:05:51,700
The new model inverts that.

1865
01:05:51,700 --> 01:05:52,900
The agent executes.

1866
01:05:52,900 --> 01:05:53,900
The human oversees.

1867
01:05:53,900 --> 01:05:55,400
The human intervenes only when needed.

1868
01:05:55,400 --> 01:05:57,000
The agent becomes the primary actor.

1869
01:05:57,000 --> 01:05:58,500
The human becomes the supervisor.

1870
01:05:58,500 --> 01:06:00,400
That shift seems subtle when you describe it.

1871
01:06:00,400 --> 01:06:02,500
In practice, it rewires everything.

1872
01:06:02,500 --> 01:06:03,600
Think about a sales team.

1873
01:06:03,600 --> 01:06:06,800
Today, a sales development rep spends their day assessing leads

1874
01:06:06,800 --> 01:06:09,000
to see who is qualified, who has the budget,

1875
01:06:09,000 --> 01:06:10,300
and who fits the target.

1876
01:06:10,300 --> 01:06:13,800
It's work that requires judgment, but happens repetitively.

1877
01:06:13,800 --> 01:06:16,300
The sales qualification agent does this assessment now.

1878
01:06:16,300 --> 01:06:17,600
The rep reviews the results.

1879
01:06:17,600 --> 01:06:19,000
The rep focuses on the leads.

1880
01:06:19,000 --> 01:06:20,800
The agent flagged as high probability.

1881
01:06:20,800 --> 01:06:22,900
They spend their time on relationships and closing

1882
01:06:22,900 --> 01:06:24,500
instead of filtering through data.

1883
01:06:24,500 --> 01:06:26,400
That's not the same job with better tools.

1884
01:06:26,400 --> 01:06:27,300
That's a different job.

1885
01:06:27,300 --> 01:06:29,000
The skills matter differently now.

1886
01:06:29,000 --> 01:06:31,200
The person who was good at quickly assessing leads

1887
01:06:31,200 --> 01:06:33,000
might not be as good at relationship building.

1888
01:06:33,000 --> 01:06:34,700
The person who was average at assessment

1889
01:06:34,700 --> 01:06:37,300
but excellent at building trust gets to spend their time

1890
01:06:37,300 --> 01:06:38,800
where they're actually strong.

1891
01:06:38,800 --> 01:06:40,200
Job descriptions change.

1892
01:06:40,200 --> 01:06:41,800
Hiring requirements change.

1893
01:06:41,800 --> 01:06:43,600
How you measure performance changes.

1894
01:06:43,600 --> 01:06:46,200
Everything flows from that one shift in what the agent does.

1895
01:06:46,200 --> 01:06:47,900
Finance teams look similar.

1896
01:06:47,900 --> 01:06:50,500
Controllers right now spend their time on reconciliation.

1897
01:06:50,500 --> 01:06:51,900
Three transactions don't match.

1898
01:06:51,900 --> 01:06:52,900
Find why.

1899
01:06:52,900 --> 01:06:54,500
Fix it manually.

1900
01:06:54,500 --> 01:06:56,600
It's detailed work that requires attention,

1901
01:06:56,600 --> 01:06:58,500
but doesn't require strategic thinking.

1902
01:06:58,500 --> 01:07:00,600
The account reconciliation agent does the matching.

1903
01:07:00,600 --> 01:07:03,000
The controller reviews exceptions and validates.

1904
01:07:03,000 --> 01:07:05,300
But the bulk of their day shifts from reconciliation

1905
01:07:05,300 --> 01:07:08,500
to analysis, to variance investigation, to strategy.

1906
01:07:08,500 --> 01:07:09,600
That's higher value work.

1907
01:07:09,600 --> 01:07:11,400
It's work that only a human should do.

1908
01:07:11,400 --> 01:07:13,600
And now a human can do it because an agent is handling

1909
01:07:13,600 --> 01:07:14,400
the baseline.

1910
01:07:14,400 --> 01:07:16,400
Security teams face the same inversion.

1911
01:07:16,400 --> 01:07:18,800
Analyst right now spend their time triaging alerts.

1912
01:07:18,800 --> 01:07:20,100
Is this a real threat?

1913
01:07:20,100 --> 01:07:21,000
Is this noise?

1914
01:07:21,000 --> 01:07:22,500
What matters?

1915
01:07:22,500 --> 01:07:24,700
The alert triage agents do this assessment.

1916
01:07:24,700 --> 01:07:26,700
The analyst reviews what the agent flagged as critical.

1917
01:07:26,700 --> 01:07:28,100
The analyst investigates.

1918
01:07:28,100 --> 01:07:30,200
The analyst makes decisions about response.

1919
01:07:30,200 --> 01:07:33,000
The analyst's time moves from filtering to investigation.

1920
01:07:33,000 --> 01:07:34,400
From assessment to action.

1921
01:07:34,400 --> 01:07:36,800
That's the job security actually wants to hire for anyway.

1922
01:07:36,800 --> 01:07:38,700
Now they can.

1923
01:07:38,700 --> 01:07:42,600
IT operations shifts to engineers handle routine troubleshooting.

1924
01:07:42,600 --> 01:07:43,800
Why is this service slow?

1925
01:07:43,800 --> 01:07:44,800
Restart it.

1926
01:07:44,800 --> 01:07:46,000
Why is this discful?

1927
01:07:46,000 --> 01:07:47,700
Clear the cache.

1928
01:07:47,700 --> 01:07:49,800
These are operational tasks that someone needs to do

1929
01:07:49,800 --> 01:07:51,700
but don't require engineering judgment.

1930
01:07:51,700 --> 01:07:53,300
Agents do these now.

1931
01:07:53,300 --> 01:07:56,200
Engineers focus on architecture, on optimization,

1932
01:07:56,200 --> 01:07:58,600
on designing systems that need less firefighting.

1933
01:07:58,600 --> 01:08:01,200
They become architects instead of on-call responders.

1934
01:08:01,200 --> 01:08:02,900
The patent is consistent across every domain.

1935
01:08:02,900 --> 01:08:05,300
Work becomes more strategic, less routine,

1936
01:08:05,300 --> 01:08:07,500
more human judgment required, less data entry,

1937
01:08:07,500 --> 01:08:10,400
less filtering, less mechanical repetition.

1938
01:08:10,400 --> 01:08:12,200
The things that made people tired and bored

1939
01:08:12,200 --> 01:08:13,200
those go to agents.

1940
01:08:13,200 --> 01:08:15,600
The things that require intuition, judgment,

1941
01:08:15,600 --> 01:08:18,200
relationship and strategy, those stay with humans.

1942
01:08:18,200 --> 01:08:20,000
And that's where the ROI actually lives.

1943
01:08:20,000 --> 01:08:21,400
It's not in replacing people.

1944
01:08:21,400 --> 01:08:23,000
It's not in doing more with fewer people.

1945
01:08:23,000 --> 01:08:25,900
It's in deploying people to work that only they can do.

1946
01:08:25,900 --> 01:08:28,400
The analyst you're paying $150,000 a year

1947
01:08:28,400 --> 01:08:31,200
is worth that investment when they're investigating threats.

1948
01:08:31,200 --> 01:08:33,500
They're not worth that investment triaging alerts.

1949
01:08:33,500 --> 01:08:35,800
Agents doing triage and analysts investigating.

1950
01:08:35,800 --> 01:08:36,800
That's the leverage.

1951
01:08:36,800 --> 01:08:39,600
This is why understanding the four layers matters organizationally,

1952
01:08:39,600 --> 01:08:40,800
not just technically.

1953
01:08:40,800 --> 01:08:42,800
Because the people implications flow directly

1954
01:08:42,800 --> 01:08:44,700
from how agents are built and governed.

1955
01:08:44,700 --> 01:08:48,300
An agent with clear ownership and audit trails is one team's can trust.

1956
01:08:48,300 --> 01:08:50,000
An agent that's orchestrated well,

1957
01:08:50,000 --> 01:08:52,900
complements human work instead of creating new problems.

1958
01:08:52,900 --> 01:08:55,400
An agent that's properly scoped does one job well

1959
01:08:55,400 --> 01:08:58,100
instead of trying to do everything and failing at all of it.

1960
01:08:58,100 --> 01:09:00,100
The shift from assistance to automation

1961
01:09:00,100 --> 01:09:01,300
isn't about technology.

1962
01:09:01,300 --> 01:09:03,000
It's about organizational design.

1963
01:09:03,000 --> 01:09:06,000
It's about what humans do when machines handle the baseline.

1964
01:09:06,000 --> 01:09:07,200
That's the actual outcome.

1965
01:09:07,200 --> 01:09:07,900
What's coming?

1966
01:09:07,900 --> 01:09:10,800
The 2026/2027 road map.

1967
01:09:10,800 --> 01:09:13,800
The platform right now is still in consolidation mode.

1968
01:09:13,800 --> 01:09:16,400
Things work, but they don't fit together smoothly yet.

1969
01:09:16,400 --> 01:09:18,900
By the end of 2026, that changes.

1970
01:09:18,900 --> 01:09:21,500
Hosted agents are approaching full general availability.

1971
01:09:21,500 --> 01:09:22,800
They're in preview now.

1972
01:09:22,800 --> 01:09:24,900
But the investment in data centers and orchestration

1973
01:09:24,900 --> 01:09:26,700
infrastructure is already committed.

1974
01:09:26,700 --> 01:09:28,400
More regions are coming online.

1975
01:09:28,400 --> 01:09:30,900
Performance is improving.

1976
01:09:30,900 --> 01:09:33,400
What started as a feature preview is becoming a standard way

1977
01:09:33,400 --> 01:09:34,800
to deploy agents at scale.

1978
01:09:34,800 --> 01:09:37,500
If you've been waiting for hosting to be stable and fully supported,

1979
01:09:37,500 --> 01:09:38,800
that window is closing.

1980
01:09:38,800 --> 01:09:43,000
By late 2026, you shouldn't think twice about hosting agents in Foundry.

1981
01:09:43,000 --> 01:09:44,400
It's the default.

1982
01:09:44,400 --> 01:09:47,000
Memory is getting smarter.

1983
01:09:47,000 --> 01:09:50,000
Right now, procedural memory exists, but it's basic.

1984
01:09:50,000 --> 01:09:52,300
The agent learns that when it encounters problem X,

1985
01:09:52,300 --> 01:09:53,400
approach Y works.

1986
01:09:53,400 --> 01:09:55,100
Useful, but limited.

1987
01:09:55,100 --> 01:09:58,200
Over the next year, procedural memory gets more sophisticated.

1988
01:09:58,200 --> 01:09:59,800
Cross-session learning is coming.

1989
01:09:59,800 --> 01:10:01,600
When an agent solves something on day one,

1990
01:10:01,600 --> 01:10:03,300
it remembers that approach on day 30.

1991
01:10:03,300 --> 01:10:04,900
Context retention improves.

1992
01:10:04,900 --> 01:10:07,100
An agent doesn't start from scratch each conversation.

1993
01:10:07,100 --> 01:10:08,900
It carries relevant context forward.

1994
01:10:08,900 --> 01:10:11,100
This matters because it reduces hallucination

1995
01:10:11,100 --> 01:10:13,100
and token consumption simultaneously.

1996
01:10:13,100 --> 01:10:15,300
An agent that remembers what it learned

1997
01:10:15,300 --> 01:10:17,800
doesn't need to retrieve the same information again.

1998
01:10:17,800 --> 01:10:19,800
Orchestration patterns are moving from preview

1999
01:10:19,800 --> 01:10:21,300
to general availability.

2000
01:10:21,300 --> 01:10:23,900
Multi-agent workflows, the stateful coordination layer,

2001
01:10:23,900 --> 01:10:24,800
is in preview.

2002
01:10:24,800 --> 01:10:25,800
That's moving to GA.

2003
01:10:25,800 --> 01:10:29,200
Group chat patterns, sequential orchestration with parallel branches,

2004
01:10:29,200 --> 01:10:32,300
handoff patterns where context flows clearly between agents.

2005
01:10:32,300 --> 01:10:33,800
These aren't experimental anymore.

2006
01:10:33,800 --> 01:10:35,200
They're supported paths.

2007
01:10:35,200 --> 01:10:37,900
That means more organizations can adopt complex workflows

2008
01:10:37,900 --> 01:10:40,800
without worrying about building on unstable foundations.

2009
01:10:40,800 --> 01:10:43,700
Integration is expanding aggressively.

2010
01:10:43,700 --> 01:10:44,700
More connectors.

2011
01:10:44,700 --> 01:10:47,400
More MCP servers registered in the ecosystem,

2012
01:10:47,400 --> 01:10:51,000
but more importantly, easier integration for organizations building their own.

2013
01:10:51,000 --> 01:10:53,900
The friction of connecting agents to custom systems is dropping.

2014
01:10:53,900 --> 01:10:56,800
By next year, integrating with your proprietary back-office system

2015
01:10:56,800 --> 01:10:58,000
should be straightforward.

2016
01:10:58,000 --> 01:11:00,200
That removes one of the barriers to scaling agents

2017
01:11:00,200 --> 01:11:02,400
across your entire technology stack.

2018
01:11:02,400 --> 01:11:05,000
Right now, agents access what's easy to access.

2019
01:11:05,000 --> 01:11:07,300
Soon, they access what matters to your business

2020
01:11:07,300 --> 01:11:09,300
regardless of how old the system is.

2021
01:11:09,300 --> 01:11:11,000
Governance is hardening.

2022
01:11:11,000 --> 01:11:14,200
Agent 365 started with discovery and identity.

2023
01:11:14,200 --> 01:11:16,500
Over the next year, more policy options arrive,

2024
01:11:16,500 --> 01:11:18,600
more granular control, better visibility

2025
01:11:18,600 --> 01:11:20,800
into what agents are accessing and why.

2026
01:11:20,800 --> 01:11:23,100
This matters because it's what lets organizations

2027
01:11:23,100 --> 01:11:24,800
deploy agents more aggressively.

2028
01:11:24,800 --> 01:11:26,800
The governance isn't a constraint anymore.

2029
01:11:26,800 --> 01:11:28,300
It's an enabler.

2030
01:11:28,300 --> 01:11:30,700
Teams can move faster because they have confidence

2031
01:11:30,700 --> 01:11:33,300
that agents are operating within appropriate bounds.

2032
01:11:33,300 --> 01:11:35,000
Cost optimization gets better tooling.

2033
01:11:35,000 --> 01:11:37,300
Right now, you measure agent costs after the fact.

2034
01:11:37,300 --> 01:11:40,600
Next year, you'll have models that predict costs before execution.

2035
01:11:40,600 --> 01:11:43,700
Cost-aware routing that picks cheaper models when quality allows.

2036
01:11:43,700 --> 01:11:46,800
Better tools for understanding where token consumption is happening.

2037
01:11:46,800 --> 01:11:48,000
For most organizations,

2038
01:11:48,000 --> 01:11:51,400
this is the difference between running one agent and running 20.

2039
01:11:51,400 --> 01:11:53,900
When you can see costs and optimize them in real time,

2040
01:11:53,900 --> 01:11:55,900
the business case changes dramatically.

2041
01:11:55,900 --> 01:11:57,900
Windows agent runtime is coming out of preview.

2042
01:11:57,900 --> 01:11:59,600
Agents running natively on Windows,

2043
01:11:59,600 --> 01:12:02,400
local execution for workloads that need to stay on premise.

2044
01:12:02,400 --> 01:12:04,700
For regulated industries, this is transformative.

2045
01:12:04,700 --> 01:12:06,500
You don't need cloud infrastructure.

2046
01:12:06,500 --> 01:12:08,900
You don't need to move data around agents run locally

2047
01:12:08,900 --> 01:12:10,600
with full access to your systems.

2048
01:12:10,600 --> 01:12:12,400
Cross-platform agents are the next step.

2049
01:12:12,400 --> 01:12:16,100
The same agent code running on Windows, Azure, and on-premises.

2050
01:12:16,100 --> 01:12:17,500
True portability.

2051
01:12:17,500 --> 01:12:19,600
Right once, deploy anywhere.

2052
01:12:19,600 --> 01:12:22,100
That's not quite here yet, but the architecture is being built.

2053
01:12:22,100 --> 01:12:23,700
By 2027, it's realistic.

2054
01:12:23,700 --> 01:12:25,900
The pattern across all of this is consistency.

2055
01:12:25,900 --> 01:12:27,500
The platform isn't getting more complex.

2056
01:12:27,500 --> 01:12:32,000
It's getting more coherent, more patterns, more frameworks, more tools.

2057
01:12:32,000 --> 01:12:34,200
But all of it building on the same foundation,

2058
01:12:34,200 --> 01:12:37,300
identity, runtime, governance, and orchestration.

2059
01:12:37,300 --> 01:12:38,700
By the end of 2026,

2060
01:12:38,700 --> 01:12:41,500
the ecosystem looks dramatically different from today,

2061
01:12:41,500 --> 01:12:43,300
not because it's fundamentally new,

2062
01:12:43,300 --> 01:12:45,800
but because it finally works as an integrated system

2063
01:12:45,800 --> 01:12:48,100
instead of a collection of separate components.

2064
01:12:48,100 --> 01:12:50,800
The structural reality, why this matters.

2065
01:12:50,800 --> 01:12:53,000
Everything we've walked through over the last hour

2066
01:12:53,000 --> 01:12:55,400
comes down to one thing, understanding the model,

2067
01:12:55,400 --> 01:12:59,400
not the products, not which agent you pick for which task, the model.

2068
01:12:59,400 --> 01:13:00,600
Because there's a shift happening

2069
01:13:00,600 --> 01:13:02,900
that doesn't make headlines the way new features do,

2070
01:13:02,900 --> 01:13:05,000
it's quieter, more structural,

2071
01:13:05,000 --> 01:13:07,700
and it determines whether your agent investments actually work.

2072
01:13:07,700 --> 01:13:10,800
The shift is from AI as a feature to AI as infrastructure.

2073
01:13:10,800 --> 01:13:12,800
Features are things you add to existing products.

2074
01:13:12,800 --> 01:13:14,400
You add copilot to your office apps.

2075
01:13:14,400 --> 01:13:17,100
It's useful, but it's still word, word is the system.

2076
01:13:17,100 --> 01:13:18,900
And copilot is just the feature bolted on.

2077
01:13:18,900 --> 01:13:20,100
Infrastructure is different.

2078
01:13:20,100 --> 01:13:22,000
Infrastructure is what systems are built on.

2079
01:13:22,000 --> 01:13:23,100
It's the foundation.

2080
01:13:23,100 --> 01:13:25,800
It's what everything else rests on and operates through.

2081
01:13:25,800 --> 01:13:28,800
Agents are transitioning from the first to the second

2082
01:13:28,800 --> 01:13:31,200
and that changes what organizations actually need to do.

2083
01:13:31,200 --> 01:13:32,200
When agents were features,

2084
01:13:32,200 --> 01:13:33,700
you didn't need to think about layers.

2085
01:13:33,700 --> 01:13:34,900
You didn't need governance.

2086
01:13:34,900 --> 01:13:37,100
You didn't need to worry about identity or orchestration.

2087
01:13:37,100 --> 01:13:38,300
You turned on copilot.

2088
01:13:38,300 --> 01:13:40,000
It worked and you were done.

2089
01:13:40,000 --> 01:13:42,200
But the moment agents start operating in your systems,

2090
01:13:42,200 --> 01:13:43,300
they stop being features.

2091
01:13:43,300 --> 01:13:44,300
They call your APIs.

2092
01:13:44,300 --> 01:13:45,400
They access your data.

2093
01:13:45,400 --> 01:13:47,200
They make decisions that affect your business.

2094
01:13:47,200 --> 01:13:48,200
They become infrastructure.

2095
01:13:48,200 --> 01:13:50,300
And infrastructure requires thinking about foundations.

2096
01:13:50,300 --> 01:13:52,400
That's what the four layers actually represent.

2097
01:13:52,400 --> 01:13:54,500
The experience layer is where humans interact.

2098
01:13:54,500 --> 01:13:55,400
That's the interface.

2099
01:13:55,400 --> 01:13:56,400
That's the front door.

2100
01:13:56,400 --> 01:13:58,700
Most organizations start here because they see a team's

2101
01:13:58,700 --> 01:14:01,300
bought or a scout agent and think that's the whole picture.

2102
01:14:01,300 --> 01:14:01,800
But it's not.

2103
01:14:01,800 --> 01:14:04,900
It's the symptom, not the system.

2104
01:14:04,900 --> 01:14:07,600
The agent layer is where domain expertise operates.

2105
01:14:07,600 --> 01:14:10,100
This is where you define what an agent actually does.

2106
01:14:10,100 --> 01:14:12,300
Security agent, sales agent, finance agent.

2107
01:14:12,300 --> 01:14:15,100
Each one specializes because specialization means accuracy.

2108
01:14:15,100 --> 01:14:17,200
But this layer by itself is just logic.

2109
01:14:17,200 --> 01:14:19,700
It can't execute anything without what comes next.

2110
01:14:19,700 --> 01:14:22,600
The runtime layer is where execution actually happens.

2111
01:14:22,600 --> 01:14:24,300
Your logic needs somewhere to run,

2112
01:14:24,300 --> 01:14:25,500
somewhere to maintain state,

2113
01:14:25,500 --> 01:14:28,200
somewhere to call tools and coordinate with other agents,

2114
01:14:28,200 --> 01:14:29,200
somewhere to scale.

2115
01:14:29,200 --> 01:14:30,300
That's the runtime.

2116
01:14:30,300 --> 01:14:32,800
It's the infrastructure the experience layer depends on

2117
01:14:32,800 --> 01:14:34,400
and the agent layer operates within.

2118
01:14:34,400 --> 01:14:37,100
The governance layer is what keeps everything safe and visible.

2119
01:14:37,100 --> 01:14:39,700
Identity, policy, audit, life cycle.

2120
01:14:39,700 --> 01:14:42,200
These aren't optional add-ons once you're operating at scale.

2121
01:14:42,200 --> 01:14:43,400
They're foundational.

2122
01:14:43,400 --> 01:14:45,500
An agent without identity is invisible.

2123
01:14:45,500 --> 01:14:47,200
An agent without policy is uncontrolled.

2124
01:14:47,200 --> 01:14:49,700
An agent without audit is unmeasurable.

2125
01:14:49,700 --> 01:14:51,600
These are the constraints that transform agents

2126
01:14:51,600 --> 01:14:55,000
from experimental features into systems organizations contrast.

2127
01:14:55,000 --> 01:14:56,100
Here's what actually matters.

2128
01:14:56,100 --> 01:14:57,800
Each layer has different concerns.

2129
01:14:57,800 --> 01:14:59,700
Each layer requires different thinking.

2130
01:14:59,700 --> 01:15:01,900
Confusing them is where most deployments break.

2131
01:15:01,900 --> 01:15:04,200
Organizations that succeed understand this.

2132
01:15:04,200 --> 01:15:06,200
They know that an agent is the experience

2133
01:15:06,200 --> 01:15:08,800
plus the logic plus the infrastructure plus the governance.

2134
01:15:08,800 --> 01:15:12,100
All four, not one, not two, all four working together.

2135
01:15:12,100 --> 01:15:14,200
Identity is the foundation because you can't govern

2136
01:15:14,200 --> 01:15:15,500
what you can't identify.

2137
01:15:15,500 --> 01:15:18,300
Entra agent ID makes agents first class principles.

2138
01:15:18,300 --> 01:15:19,300
That sounds technical.

2139
01:15:19,300 --> 01:15:22,000
What it actually means is agents become manageable.

2140
01:15:22,000 --> 01:15:22,700
You know what they are.

2141
01:15:22,700 --> 01:15:24,000
You can control what they do.

2142
01:15:24,000 --> 01:15:25,400
You can audit what they did.

2143
01:15:25,400 --> 01:15:28,500
Orchestration is the multiplier because a single agent is useful.

2144
01:15:28,500 --> 01:15:31,000
But multiple agents coordinating are transformative.

2145
01:15:31,000 --> 01:15:34,600
But coordination only works if you have infrastructure designed for it.

2146
01:15:34,600 --> 01:15:35,600
That's the runtime layer.

2147
01:15:35,600 --> 01:15:37,700
Foundry isn't just a place to host agents.

2148
01:15:37,700 --> 01:15:39,500
It's where orchestration becomes possible.

2149
01:15:39,500 --> 01:15:41,300
Understanding this model isn't optional.

2150
01:15:41,300 --> 01:15:43,800
It's what separates organizations building agent infrastructure

2151
01:15:43,800 --> 01:15:46,500
from organizations experimenting with agent features.

2152
01:15:46,500 --> 01:15:48,100
One scales, the other doesn't.

2153
01:15:48,100 --> 01:15:50,200
The real question isn't which agent you should use.

2154
01:15:50,200 --> 01:15:51,800
It's what your agent architecture looks like.

2155
01:15:51,800 --> 01:15:53,400
You're not choosing between products.

2156
01:15:53,400 --> 01:15:56,100
You're choosing how AI operates in your organization.

2157
01:15:56,100 --> 01:15:58,800
Start with one agent, ground it in your data,

2158
01:15:58,800 --> 01:16:02,800
register it in governance, measure what it does, then scale.

2159
01:16:02,800 --> 01:16:05,200
The teams that understand the four layers move faster

2160
01:16:05,200 --> 01:16:06,500
and make better decisions.

2161
01:16:06,500 --> 01:16:08,600
Everything else is just picking the right tool

2162
01:16:08,600 --> 01:16:09,800
for the layer you're building.