June 15, 2026

Building Multi-Agent AI Systems with Copilot Studio: From Ideas to Intelligent Automation with David Lorenzo Lopez [MVP]

Multi-Agent Systems, Copilot Studio, Microsoft Agent Framework, Intelligent Automation, AI Orchestration, Microsoft Foundry, Artificial Intelligence, David Lorenzo Lopez, MC65 Podcast, Generative AI, Autonomous Agents, Enterprise AI, Microsoft Azure, Microservices Architecture, LLM Agents, AI for Business, GPT-3.5, Software Development, Microsoft Ecosystem

You can transform your business with Microsoft Copilot Studio by building modular and scalable Multi-Agent AI Systems. The platform’s low-code design lets you create and manage ai agents, whether you are just starting or already an expert. Microsoft Copilot Studio helps you automate tasks, reduce costs, and speed up response times—some teams have cut manual work by 75% and improved customer satisfaction. Responsible ai practices, like human-in-the-loop, ensure you keep control. Seamless integration with Microsoft’s ecosystem supports your journey to smarter automation.

Key Takeaways

Microsoft Copilot Studio allows you to build modular and scalable multi-agent AI systems with a low-code approach.
Multi-agent AI systems enhance efficiency by automating tasks, which can reduce manual work by up to 75%.
Define clear roles for each agent to ensure they focus on specific tasks and improve overall workflow.
Utilize the Model Context Protocol (MCP) to enhance communication between agents, ensuring smooth task transitions.
Regularly test and debug your agents to catch issues early and maintain a high level of service.
Integrate agents with Microsoft 365 and Azure services to expand their capabilities and streamline operations.
Monitor agent performance using analytics tools to identify trends and improve effectiveness over time.
Stay engaged with the Copilot Studio community for support, resources, and best practices to enhance your skills.

Multi-Agent AI Systems Overview

What Are Multi-Agent AI Systems

You encounter multi-agent ai systems in many modern organizations. These systems use multiple autonomous agents that work together to solve complex problems. Each agent operates independently, focusing on specific tasks and sharing information with others. You see decentralization as a key feature, where agents use local data and make decisions without a central authority. Agents interact socially, negotiate, and cooperate to reach shared goals. They respond proactively to changes and communicate through structured channels.

Characteristic	Description
Decentralization	Agents act independently using local data, aiming for a common goal.
Local Views	Each agent sees only what it needs, not the whole system.
Social Interaction and Coordination	Agents negotiate, cooperate, and resolve conflicts for teamwork.
Proactivity	Agents predict needs and respond to changes in the environment.
Communication Channels	Agents use structured protocols to share information and ensure smooth operations.

Multi-agent ai systems help you break down large tasks into smaller, manageable parts. You gain flexibility and adaptability, which are essential for business growth.

Business Value and Use Cases

You benefit from multi-agent ai systems in many business areas. These systems automate processes, improve accuracy, and speed up operations. You can use them in finance, customer service, and supply chain management. Specialized agents handle tasks like purchase requests, vendor management, and payment scheduling. In customer experience, agents manage onboarding, service requests, and proactive support. Supply chain operations use agents for demand forecasting, inventory monitoring, and replenishment.

Business Area	Use Case Description
Finance and Accounting	Automating procure-to-pay processes with agents for purchase requests, vendor management, and payment scheduling.
Customer Experience	Managing customer journeys with onboarding, service requests, and proactive support through specialized agents.
Supply Chain Operations	Optimizing inventory management with forecasting, monitoring, and replenishment agents.

You see measurable improvements in efficiency and customer satisfaction when you use multi-agent ai systems.

Microsoft Copilot Studio’s Role

Microsoft Copilot Studio gives you the tools to design, deploy, and monitor multi-agent ai systems. You define task-specific agents that manage different parts of your workflow. These agents communicate, share results, and handle exceptions, making task sharing seamless across platforms.

You orchestrate interconnected agents for effective collaboration.
You enhance modularity and scalability by assigning agents to specific tasks.
You enable agents to communicate and manage exceptions for smooth operations.

Microsoft Copilot Studio has moved from single-agent models to multi-agent orchestration layers. You use Agent-to-Agent protocols for communication, which supports scalability in complex workflows. The focus on orchestration helps you manage enterprise-level tasks efficiently. You build systems that grow with your business needs and adapt to new challenges.

Tip: Start with small workflows and scale up as you gain confidence in managing multi-agent ai systems. Microsoft Copilot Studio supports both low-code and pro-code development, so you can tailor solutions to your skill level.

Workspace Setup in Copilot Studio

Setting up your workspace in Copilot Studio is the first step to building effective multi-agent AI systems. You will find that the platform guides you through each stage, making it easy to start, organize, and optimize your projects.

Creating a Project

You begin by creating a new copilot project. Copilot Studio uses a step-by-step process that helps you define your goals and set up your workspace. Here is a typical workflow:

Go to Copilot Studio and start a new copilot.
Describe your copilot’s purpose using natural language.
Define how your copilot should respond to users.
Choose web sources for your copilot’s answers.
Start the creation process and wait for configuration.
Remove example training topics to keep your workspace clean.
Add Power Apps (dataverse) as a data source.
Select the relevant tables from dataverse.
Define the purpose of each data source.
Add synonyms and glossary terms for better understanding.
Wait for indexing to finish.
Enable the general knowledge feature.
Choose between Classic or Generative interaction modes.
Test your copilot with websites and dataverse data.
Deploy your copilot to Teams.
Run example questions to check for issues.

Tip: Test your copilot early and often. This helps you catch issues before they affect users.

Navigating the Interface

Copilot Studio’s interface gives you the tools you need to build ai agents and manage your workspace efficiently. You will see several key features:

Feature	Description
Build conversational capabilities	Design agents that respond to natural language prompts and guide workflows.
Add autonomous capabilities	Create agents that manage tasks on their own and escalate when needed.
Use pre-built agents and templates	Access ready-to-use agents or customize templates for your needs.
Create multi-agent systems	Orchestrate complex processes by routing tasks to the right agent.

You can also use agent flows to automate processes, write natural language instructions, and integrate AI tools for specific tasks. The interface makes it easy to switch between these features, so you can focus on what matters most.

Note: Regularly monitor your agents and connections. This helps you avoid issues like dormant agents or misconfigurations.

Model Selection

Choosing the right AI model for your agents is important. Copilot Studio offers several model types, each designed for specific tasks:

Tag	Description	Strengths	Latency	Cost	Reasoning depth
Deep	For complex, multi-step reasoning and workflows	Analytics, troubleshooting, policy analysis	Highest	Highest	Multi-step, tool-rich
Auto	For mixed workloads and dynamic routing	Helpdesk, employee agents, blended tasks	Variable	Variable	Multi-step, tool-rich
General	For fast, cost-effective everyday chat and simple automation	Drafting, summarizing, FAQs, light actions	Lowest	Lowest	Shallow-to-moderate

Selecting the right model ensures your agents perform well and stay within budget. For example, use a Deep model for complex analysis or a General model for quick replies. Always match the model to your business needs.

Remember: The right model improves user satisfaction and helps you manage costs.

Agent Roles & Planning

Defining Agent Roles

You start your project planning by defining clear roles for your agents. Each agent should have a specific purpose that matches your business needs. In Copilot Studio, you can use both low-code and pro-code approaches to build agents. You decide if you need child agents for simple tasks or connected agents for more complex workflows. This choice helps you coordinate multiple agents and keep your system organized.

When you define agent roles, you should:

Identify the main functions in your business, such as HR, IT, sales, or website automation.
Assign agents to handle these functions, making sure each agent has a clear job.
Break down your workflows into smaller stages. This lets you see where agents can make the biggest impact.
Focus on tasks that happen often and take a lot of effort. Avoid giving agents tasks that are too broad or too narrow.
Make sure you follow security and governance best practices for every agent you deploy.
Use integration options with Microsoft 365, Azure AI Foundry, and Fabric to connect your agents to the right data and tools.

A good way to define agent roles is to look for Core Units of Work. These are tasks that you can measure and that help your business reach its goals. By focusing on these units, you make your planning more effective and your agents more useful.

Tip: Review your agent roles regularly. As your business changes, you may need to update your planning and adjust agent responsibilities.

Task Automation

You can automate your planning by letting agents handle repetitive or complex tasks. In Copilot Studio, agents can do much more than just answer questions. They can retrieve data from different systems, update records, send emails, and trigger workflows. You can also connect agents to external platforms, making your automation even more powerful.

Here are some ways agents automate tasks:

Pull customer support ticket information from Salesforce and create personalized responses.
Query live databases using Dataverse to solve user requests right away.
Trigger automated workflows, such as processing invoice requests and sending them through Teams or Outlook.
Connect to APIs, data sources, and enterprise systems for seamless operations.
Build custom agents for specific roles, workflows, or tasks using low-code tools.
Create autonomous agents that can plan, act, and adapt with little human input.

When you automate your planning, you save time and reduce errors. Agents can work around the clock, making sure your business runs smoothly. You also free up your team to focus on more important work.

Note: Always test your automated workflows before you go live. This helps you catch problems early and keeps your operations running smoothly.

Collaboration Patterns

Planning for collaboration is key when you use multiple agents in your AI system. Each agent works on its own tasks, but they also need to share information and coordinate actions. This teamwork mirrors how people work together in a business.

Multi-agent orchestration lets you coordinate multiple agents to solve problems. Each agent focuses on a specific task, which makes your system easier to manage and scale. You can use communication protocols so agents can exchange information, assign responsibilities, and coordinate actions. This includes breaking down work, sharing resources, and resolving conflicts.

Think of collaboration like a fleet of drones searching a disaster site. Each drone works on its own, but they also share what they find and adjust their actions based on new information. In your business, agents do the same. They work independently but also communicate and adapt to changes.

AI-Optimized API Architecture helps your agents work together in real time. This setup allows agents to make decisions and act as a team, which is important for enterprise applications.

Tip: Use clear communication strategies in your planning. Make sure your agents know when and how to share information for the best results.

A strong project planning process includes defining agent roles, automating tasks, and setting up effective collaboration patterns. When you plan well, your agents can deliver real value to your business.

Orchestration & Integration

Multi-Agent Orchestration

You orchestrate agents in Copilot Studio by coordinating their actions and responsibilities. You can use different strategies to manage how agents work together. Inline agents act as small, reusable workflows inside a main agent. These agents share context, which makes passing data simple. You keep inline agents focused on one task and test them often to ensure reliability.

Connected agents operate as separate units. Each agent has its own tools and knowledge. The main agent delegates tasks to connected agents, which helps you achieve modularity and domain separation. You decide when to hand off tasks based on clear criteria. You manage context and parameters passed between agents to maintain accuracy. Security remains important, so you make sure connected agents follow restrictions. You log and monitor sessions to track performance and debug issues.

You create separate agents only when a subtask is complex, needs different governance rules, or can be reused across multiple main agents. This approach helps you automate complex workflows and scale your system as your business grows.

Tip: Plan your orchestration carefully. You improve modularity and scalability by assigning agents to specific tasks and monitoring their performance.

Model Context Protocol (MCP)

You use the Model Context Protocol to enhance communication between agents. MCP structures the way agents share context, including user intent, previous actions, and the current state of the system. This protocol helps agents coordinate and transition tasks smoothly. You ensure that agents understand what has happened and what needs to happen next.

MCP supports seamless collaboration. Agents use MCP to exchange information, which reduces confusion and improves accuracy. You rely on MCP to keep workflows efficient and responsive. When agents share context, you avoid errors and maintain a high level of service.

Note: MCP helps you connect to the right data and maintain consistency across your multi-agent AI system.

API & Azure Integration

You integrate agents with APIs and Azure services to expand their capabilities. Copilot Studio connects easily to Microsoft Power Platform connectors, which lets you access business data and automate tasks. You select data sources that matter most to your business. You set up Data Loss Prevention (DLP) policies to protect sensitive information.

You deploy agents in key regions and high-impact teams to showcase their value. You invest time in planning, data source selection, and security to design effective agents. You measure agent success using metrics like engagement rates and customer satisfaction scores.

Copilot Studio agents work with Fabric agents to reason over enterprise data at scale. This improves accuracy and relevance. Multi-agent orchestration allows specialized agents to manage their expertise while providing cohesive responses to users. You automate complex workflows and achieve seamless automation across your organization.

Tip: Integrate agents with Azure and APIs to unlock advanced features and streamline your business operations.

Integration Feature	Benefit
Power Platform connectors	Access business data and automate tasks
DLP policies	Protect sensitive information
Azure services	Enhance agent capabilities
Fabric agents collaboration	Improve accuracy and relevance

You build a robust AI system by connecting agents, using MCP, and integrating with Azure and business data. You achieve modularity, scalability, and seamless workflow automation.

Agent Configuration & Testing

Setting Parameters

You need to configure your agents before they can work in your ai system. Start by setting clear parameters for each agent. These parameters control how agents behave, what data they use, and how they interact with users. You can adjust things like response style, allowed actions, and access to business data. For example, you might set an agent to answer only customer service questions or limit its access to sensitive information.

You should also define triggers and conditions. These settings tell agents when to start a task or when to ask for help from a human. By setting these rules, you make sure your ai system stays safe and reliable. You can use Copilot Studio’s interface to update parameters quickly as your needs change.

Tip: Review your agent parameters often. This helps you keep your ai system secure and up to date.

Simulation & Debugging

Testing your agents is an important step in building a strong ai system. You can use Copilot Studio’s simulation tools to see how agents respond to real-world situations. Run sample conversations and workflows to check if agents understand instructions and complete tasks correctly.

If you find problems, use debugging tools to trace what happened. Look at logs and error messages to find out where things went wrong. You can fix issues by changing parameters or updating agent logic. Repeat your tests until agents work as expected.

Run simulations for different scenarios.
Check agent responses for accuracy.
Use logs to find and fix errors.

Note: Testing helps you catch mistakes early. This saves time and builds trust in your ai system.

Iteration & Versioning

You will improve your agents over time. Each update makes your ai system smarter and more reliable. To manage these changes, you need a good versioning strategy. Keep a living document that explains each agent’s design, purpose, and updates. This helps everyone understand how agents work and why changes happen.

Use Application Lifecycle Management (ALM) features in Copilot Studio to export and save agent configurations. Store these files in a source repository so you can track changes and roll back if needed. You can also set up Continuous Integration and Continuous Deployment (CI/CD) processes. These tools automate testing and deployment, making sure every update is safe and well-documented.

Here is a table that shows three effective strategies for managing agent versions:

Strategy	Description
Documentation	Keep a living document with agent design, purpose, and changes.
Version Control	Use ALM features to export and version agent configurations in a source repository.
CI/CD Integration	Automate export/import and testing of agent solutions for safe and reliable updates.

Tip: Good versioning helps you grow your ai system with confidence. You can always see what changed and why.

Responsible AI & Governance

Human-in-the-Loop

You play a key role in responsible AI by keeping humans involved in important decisions. In multi-agent AI systems, you can set up checkpoints where agents ask for your approval before taking sensitive actions. This approach helps you maintain control and traceability. You should also assign clear roles to each agent so everyone knows their responsibilities. When agents communicate and share results, you can review their work and step in if needed.

Here are some best practices for human-in-the-loop mechanisms:

Best Practice	Description
Secure-by-design principles	Enforce least-privilege access, segregate duties, and keep detailed logs.
Human-in-the-loop review	Require explicit approvals for sensitive actions to ensure traceability.
Clear role assignments	Define agent responsibilities for better collaboration and workflow.

You can also use orchestration patterns where agents pass tasks and outputs to each other. This teamwork lets you focus on reviewing only the most critical steps.

Tip: Set up checkpoints for human review in workflows that handle personal or financial data.

Security & Compliance

You must protect your organization’s data and follow strict compliance standards when building AI systems. Copilot Studio supports secure development by following the Software Development Lifecycle (SDL). This includes encrypting data at rest and in transit, keeping production and test environments separate, and documenting secure coding practices. You also need to complete threat modeling and keep audit logs for all agent interactions.

Here is a table of key security and compliance assessments:

Assessment Type	Key Components
Software Development Lifecycle	Service metadata, no production data in test, encryption, threat modeling, secure standards, logging
Accessibility Tests	Aligned to Microsoft’s accessibility standards
Responsible AI Evaluations	Expert reviews for accuracy, consistency, and inclusivity
Tenant Trust Evaluations	Security questionnaires, IT council reviews, detailed documentation
Works Council Evaluations	Privacy standards for personal data and employee monitoring

You should always control data access and tool permissions for each agent. Segment knowledge by role to reduce risk and limit exposure. Make sure agents only retrieve data from approved sources with proper filtering. Copilot Studio uses Power Platform governance features, such as data loss prevention, compliance certifications, geographic data residency, and audit logging through Microsoft Purview.

Note: Protecting your data and following compliance rules builds trust with users and regulators.

Monitoring & Analytics

You need to track how your agents perform and make sure they follow compliance rules. Copilot Studio gives you several tools for monitoring and analytics. You can use built-in dashboards or connect to external tools for deeper insights.

Here are some ways to monitor and analyze your AI system:

Use evaluation frameworks like Microsoft Conversational AI Quality (CAIQ) to measure performance.
Track key metrics such as intent recognition accuracy, response relevance, resolution rate, fallback rate, user engagement, and goal completion rate.
Access analytics dashboards in Copilot Studio, Power BI, Azure Application Insights, or Bot Framework Analytics to review agent activity.

You should review these metrics often to spot trends and fix problems quickly. Monitoring helps you improve agent performance and ensures your system handles data responsibly.

Tip: Set up alerts for unusual activity so you can respond to issues before they affect users.

Optimization & Resources

Performance Tuning

You can make your multi-agent AI system work better by tuning its performance in Copilot Studio. Start by understanding what each agent should do and where it might have limits. When you know the agent’s scope, you can set clear goals and measure progress.

You do not need to write code to improve your agents. Copilot Studio lets you use no-code tools to customize and tune agents for your needs. You can create agents for specific tasks, which helps your system run faster and more accurately. When you connect your agents to Microsoft 365 data, you make sure they always use the latest information.

Here are some ways to tune performance:

Use task-specific agents for important jobs. This improves speed and consistency.
Test your agents with real scenarios. You can measure quality, usability, and how fast they respond.
Compare different versions of your agents side by side. This helps you spot problems before you update your system.
Grade your agents across a full set of tests. This gives you a clear picture of how well your system works.
Write clear and specific prompts for your agents. This guides them to give better answers.
Ask users for feedback. Their ideas help you improve agent quality over time.

Tip: Review your agents often. Small changes can make a big difference in how well your system works.

Community & Support

You do not have to solve problems alone. The Copilot Studio community gives you many ways to get help and share ideas. You can join the Microsoft Community Hub for Copilot Studio to ask questions and learn from others. You can also find step-by-step guides and best practices in the official documentation.

If you need more support, you can reach out to UCOP IT Services for Copilot Studio. These resources help you solve problems quickly and keep your projects on track.

Resource Type	Where to Find It
Community Discussions	Microsoft Community Hub for Copilot Studio
How-to Guides	Official Copilot Studio Documentation
IT Support	UCOP IT Services for Copilot Studio

Note: Joining the community helps you stay updated and learn new tips from other users.

Further Learning

You can keep building your skills with many learning resources. Start by exploring how to define and build AI agents in Copilot Studio. Learn the difference between child agents and connected agents, and when to use each. Discover how to orchestrate agents for different business areas like HR, IT, and sales. You can also learn how to connect your agents to Microsoft 365, Azure AI Foundry, and Fabric.

To master security and deployment, check out best practices for governance and quick deployment to Teams, SharePoint, and web apps. You can find these topics in the Microsoft Copilot Studio documentation, Mastering Copilot Studio video series, and the Copilot Studio Agent Academy. You can also join Agent in a Day events for hands-on practice.

Here are some helpful links:

Tip: Keep learning and practicing. The more you explore, the better your multi-agent AI systems will become.

You have learned how to build multi-agent AI systems with Copilot Studio. This platform gives you many benefits:

Improved performance with faster agent responses
Simple automation and decision-making tools
Easy integration with business data sources
Better team collaboration through shared profiles and skills

Organizations now use Copilot Studio to create agents that delegate tasks, making workflows more efficient and complex tasks easier to manage.

To grow your skills, explore courses on AI agents, data visualization, and advanced agent building. Stay updated with new features and best practices in Copilot Studio.

FAQ

How do you start building a multi-agent AI system in Copilot Studio?

You begin by creating a new project. Define your goals, select agent roles, and choose the right AI models. Copilot Studio guides you through each step with a simple interface.

Can you use Copilot Studio without coding experience?

Yes, you can build agents using low-code tools. The platform offers templates and drag-and-drop features. You do not need to write code to automate tasks or workflows.

What types of business data can you connect to Copilot Studio agents?

You connect agents to Microsoft 365, Dataverse, Azure, and other business data sources. Agents retrieve, update, and process information from these platforms to automate tasks.

How do you ensure your AI agents follow security and compliance rules?

You set permissions, use data loss prevention policies, and monitor agent activity. Copilot Studio supports encryption and audit logging. You control access to sensitive data.

What is the Model Context Protocol (MCP) and why does it matter?

MCP helps agents share context, such as user intent and workflow state. This protocol improves coordination and accuracy. You use MCP to keep multi-agent systems efficient.

How can you test and improve your agents in Copilot Studio?

You run simulations, check agent responses, and use debugging tools. Copilot Studio lets you update parameters and version agents. You track performance with built-in analytics.

Where can you find support and learning resources for Copilot Studio?

You join the Microsoft Community Hub, read official documentation, and watch video tutorials. You also attend Agent Academy events for hands-on practice.

Tip: Explore the Copilot Studio learning hub for step-by-step guides and best practices.

🚀 Want to be part of m365.fm?

Then stop just listening… and start showing up.

👉 Connect with me on LinkedIn and let’s make something happen:

🎙️ Be a podcast guest and share your story
🎧 Host your own episode (yes, seriously)
💡 Pitch topics the community actually wants to hear
🌍 Build your personal brand in the Microsoft 365 space

This isn’t just a podcast — it’s a platform for people who take action.

🔥 Most people wait. The best ones don’t.

👉 Connect with me on LinkedIn and send me a message:
"I want in"

Let’s build something awesome 👊

1
00:00:00,000 --> 00:00:04,000
Welcome to another edition of the MC65 podcast.

2
00:00:04,000 --> 00:00:07,000
Today I'm joined by David Lorenzo Lopez,

3
00:00:07,000 --> 00:00:10,240
a technology leader, community contributor mentor

4
00:00:10,240 --> 00:00:13,680
and passionate advocate for intelligent automation.

5
00:00:13,680 --> 00:00:16,600
Since 2010 David has helped organizations

6
00:00:16,600 --> 00:00:18,160
such as Pauline Business Genitals

7
00:00:18,160 --> 00:00:21,160
into practical solution, solutions,

8
00:00:21,160 --> 00:00:23,080
spanning, enterprising, development,

9
00:00:23,080 --> 00:00:25,000
cloud technologies, and recently,

10
00:00:25,000 --> 00:00:26,960
iPod waste of working.

11
00:00:26,960 --> 00:00:29,880
Over the last years, his focus has shifted

12
00:00:29,880 --> 00:00:32,640
heavily towards intelligent automation,

13
00:00:32,640 --> 00:00:35,240
co-pilot studio, Microsoft Foundry Azure,

14
00:00:35,240 --> 00:00:37,800
and building AI systems at the level of unmeasurable

15
00:00:37,800 --> 00:00:38,920
business outcomes.

16
00:00:38,920 --> 00:00:41,200
Behind this professional work, David is well known

17
00:00:41,200 --> 00:00:44,560
for as community speaker, who regularly shares knowledge

18
00:00:44,560 --> 00:00:48,000
across the globe and helps us mentor,

19
00:00:48,000 --> 00:00:51,480
so yeah, the generation of technology professors.

20
00:00:51,480 --> 00:00:54,520
In this episode, we are diving deep into one

21
00:00:54,520 --> 00:00:57,240
of the hottest topics in AI right now,

22
00:00:57,240 --> 00:01:00,880
creating all end-or-casturing multi-agent systems

23
00:01:00,880 --> 00:01:03,200
using co-pilot studio, Microsoft Foundry,

24
00:01:03,200 --> 00:01:06,480
and Microsoft agent framework.

25
00:01:06,480 --> 00:01:08,720
David, welcome to the show.

26
00:01:08,720 --> 00:01:09,960
- Thanks, Michael.

27
00:01:09,960 --> 00:01:11,400
It's a pleasure to be here.

28
00:01:11,400 --> 00:01:12,640
Thanks for inviting me.

29
00:01:12,640 --> 00:01:17,480
Really, yeah, really looking forward to our talk today.

30
00:01:17,480 --> 00:01:18,640
- Yeah, thank you.

31
00:01:18,640 --> 00:01:24,560
Can you tell a little bit about your background

32
00:01:24,560 --> 00:01:29,560
and how you join the technology or the Microsoft ecosystem?

33
00:01:29,560 --> 00:01:32,640
- Yeah, of course.

34
00:01:32,640 --> 00:01:37,640
So I started back in 2008, my, let's say,

35
00:01:37,640 --> 00:01:41,240
my programming journey.

36
00:01:41,240 --> 00:01:47,360
I started doing personal projects

37
00:01:47,360 --> 00:01:52,200
after my high school studies,

38
00:01:52,200 --> 00:01:57,200
and in parallel, I take up the computer science studies

39
00:01:57,200 --> 00:02:01,800
in the university.

40
00:02:01,800 --> 00:02:06,800
So then I started programming in PHP back in the days,

41
00:02:06,800 --> 00:02:14,160
making WordPress, sites, and yeah, personal projects,

42
00:02:14,160 --> 00:02:15,200
press and our websites.

43
00:02:15,200 --> 00:02:17,160
Like, usually back in the days,

44
00:02:17,160 --> 00:02:20,720
was one of the most common languages

45
00:02:20,720 --> 00:02:24,960
to learn programming in the website.

46
00:02:24,960 --> 00:02:29,960
And then in 2010, I started working in companies

47
00:02:29,960 --> 00:02:36,320
and in my first company,

48
00:02:36,320 --> 00:02:39,440
they were programming in DocNet.

49
00:02:39,440 --> 00:02:43,240
It was a desktop application,

50
00:02:43,240 --> 00:02:46,200
but yeah, I like Shisharp,

51
00:02:46,200 --> 00:02:49,400
I like all the SQL server stuff,

52
00:02:49,400 --> 00:02:51,600
and then from there,

53
00:02:51,600 --> 00:02:55,760
I started my journey inside the Microsoft ecosystem.

54
00:02:55,760 --> 00:02:58,600
So yeah, my background in the first days,

55
00:02:58,600 --> 00:03:02,000
was all about web development.

56
00:03:02,000 --> 00:03:05,600
Even Azure didn't exist back in the days.

57
00:03:05,600 --> 00:03:10,600
So we used to publish in Internet Information Server

58
00:03:10,600 --> 00:03:16,480
and also using ASP.NET and, yeah, the web forms,

59
00:03:19,240 --> 00:03:23,280
even I did some projects in the web forms.

60
00:03:23,280 --> 00:03:28,200
Now you are deep in the AI part.

61
00:03:28,200 --> 00:03:31,240
So what's there a moment,

62
00:03:31,240 --> 00:03:32,440
called the wow moment,

63
00:03:32,440 --> 00:03:35,280
where you say this change everything?

64
00:03:35,280 --> 00:03:36,680
Yeah, yeah, of course.

65
00:03:36,680 --> 00:03:41,680
So I started with the first GPD that comes to the public,

66
00:03:41,680 --> 00:03:47,040
it was a GPD three.

67
00:03:48,320 --> 00:03:49,840
Of course, before that,

68
00:03:49,840 --> 00:03:54,840
I tried to create some bots using the bot framework,

69
00:03:54,840 --> 00:03:57,200
using PowerBit to Allegiance,

70
00:03:57,200 --> 00:03:59,680
the old copy studio, let's say.

71
00:03:59,680 --> 00:04:04,200
But of course, it was difficult to have

72
00:04:04,200 --> 00:04:05,880
something that really works well

73
00:04:05,880 --> 00:04:10,840
because you need to guide the conversation flow too much,

74
00:04:10,840 --> 00:04:13,560
putting a lot of guardrails,

75
00:04:13,560 --> 00:04:18,560
and of course, if the user was out of these guardrails,

76
00:04:18,560 --> 00:04:23,560
the customer was lost in the conversation.

77
00:04:23,560 --> 00:04:28,560
So after that, I started trying GPD three.

78
00:04:28,560 --> 00:04:33,680
five and I really say, wow, this is a really game changer

79
00:04:33,680 --> 00:04:39,760
because you can talk of almost everything

80
00:04:39,840 --> 00:04:44,840
with the chatbot, you can also restore the conversation

81
00:04:44,840 --> 00:04:48,800
or guide the conversation,

82
00:04:48,800 --> 00:04:50,600
you don't lose the customer.

83
00:04:50,600 --> 00:04:53,680
So it was really a game changing.

84
00:04:53,680 --> 00:04:55,760
Then of course, after that,

85
00:04:55,760 --> 00:05:02,120
what we are working now is to convert that

86
00:05:02,120 --> 00:05:06,720
into value for the companies.

87
00:05:06,720 --> 00:05:08,320
And that is not being easy.

88
00:05:08,320 --> 00:05:12,600
We can talk about this in further in the episode,

89
00:05:12,600 --> 00:05:17,600
but one thing is all the demos, all the perfect pixel demos

90
00:05:17,600 --> 00:05:24,800
we are seeing there in the events or in YouTube videos.

91
00:05:24,800 --> 00:05:29,000
And then when you try to do real value

92
00:05:29,000 --> 00:05:31,280
for the customer, for the companies,

93
00:05:31,280 --> 00:05:34,160
the reality is much more different.

94
00:05:34,160 --> 00:05:35,920
It's not so easy to do it.

95
00:05:35,920 --> 00:05:40,160
So yeah, now I am in that part,

96
00:05:40,160 --> 00:05:44,280
fighting to put the feeds in the floor

97
00:05:44,280 --> 00:05:49,280
because people are flying in the clouds with AI

98
00:05:49,280 --> 00:05:53,840
and then the reality is that you need much more effort

99
00:05:53,840 --> 00:05:57,680
that what is looking in the, in the,

100
00:05:57,680 --> 00:06:00,680
in the videos in the demos that you see

101
00:06:00,680 --> 00:06:01,880
out there in the internet.

102
00:06:04,680 --> 00:06:09,680
They, they, you are well known for multi agent AI systems.

103
00:06:09,680 --> 00:06:13,800
For this, who are not, yeah, familiar with the concept,

104
00:06:13,800 --> 00:06:17,160
what exactly is the multi agent system?

105
00:06:17,160 --> 00:06:20,400
- Yeah, so multi agents are,

106
00:06:20,400 --> 00:06:25,400
I put some, sometimes comparison for those

107
00:06:25,400 --> 00:06:28,960
who come from web development or let's say,

108
00:06:28,960 --> 00:06:30,640
application development.

109
00:06:30,640 --> 00:06:34,280
We had in the beginning, we had a monolith application

110
00:06:34,280 --> 00:06:38,880
where we developed everything inside one project.

111
00:06:38,880 --> 00:06:44,200
So if we have, let's put the sample of e-commerce,

112
00:06:44,200 --> 00:06:47,280
we have the shopping card, we have the orders,

113
00:06:47,280 --> 00:06:50,240
we have the products catalog.

114
00:06:50,240 --> 00:06:55,240
So we used to create one project with several connections

115
00:06:55,240 --> 00:06:59,360
to the database, to get the product catalog,

116
00:06:59,360 --> 00:07:03,120
to filter out, to add products to the shopping card

117
00:07:03,120 --> 00:07:05,720
to create the orders, to cancel the orders.

118
00:07:05,720 --> 00:07:08,440
All that comes in a monolith project.

119
00:07:08,440 --> 00:07:13,440
We pull this that to the IAS server and then we are out.

120
00:07:13,440 --> 00:07:18,440
What happens then, if we want to test one specific part

121
00:07:18,440 --> 00:07:22,080
of the project, we have troubles because we want to,

122
00:07:22,080 --> 00:07:25,840
we need to test everything in a global test.

123
00:07:25,840 --> 00:07:32,160
So after that, we learn to split in what we call microservices.

124
00:07:32,840 --> 00:07:36,040
So then we have the shopping card microservice,

125
00:07:36,040 --> 00:07:38,520
the order management microservice,

126
00:07:38,520 --> 00:07:40,760
the product catalog microservice,

127
00:07:40,760 --> 00:07:45,240
and all that was orchestrator with a website

128
00:07:45,240 --> 00:07:49,760
or Kubernetes or whatever technology we use.

129
00:07:49,760 --> 00:07:54,560
So if we translate that concept to agents,

130
00:07:54,560 --> 00:07:59,120
we started back in 2023, 2024,

131
00:07:59,120 --> 00:08:03,720
creating a big agent with a big system prompt,

132
00:08:03,720 --> 00:08:08,080
with a bunch of tools or functions or plugins

133
00:08:08,080 --> 00:08:11,800
or whatever name you want to put in front.

134
00:08:11,800 --> 00:08:17,000
And then, okay, now the agent can make calculations

135
00:08:17,000 --> 00:08:21,880
or do math operations, or it can query my API.

136
00:08:21,880 --> 00:08:23,560
Wow, that was incredible.

137
00:08:23,560 --> 00:08:27,640
But then when we added three, four, five APIs,

138
00:08:27,640 --> 00:08:32,200
what happens is the agent get confused

139
00:08:32,200 --> 00:08:36,600
with which tool should I call or now,

140
00:08:36,600 --> 00:08:40,200
or if we introduce the concept of dynamic training,

141
00:08:40,200 --> 00:08:44,080
we need to call two, three tools in a row

142
00:08:44,080 --> 00:08:48,120
for creating a response for the user query,

143
00:08:48,120 --> 00:08:50,760
then the agent will mess up with that.

144
00:08:50,760 --> 00:08:55,800
So we learn also to split that bunch of tools,

145
00:08:55,800 --> 00:09:03,080
skills or actions that the agent can do in specific agents.

146
00:09:03,080 --> 00:09:05,960
So let's compare the agent to the microservice.

147
00:09:05,960 --> 00:09:14,000
So every agent has a specific, let's say, functionality,

148
00:09:14,000 --> 00:09:24,680
specific tools to be able to respond or take some actions.

149
00:09:24,680 --> 00:09:31,280
And then we create small agents that then can be called

150
00:09:31,280 --> 00:09:36,720
from an orchestrator agent to create different actions,

151
00:09:36,720 --> 00:09:40,440
the different responses that the user needs.

152
00:09:40,440 --> 00:09:46,360
So if we come again with the shopping, the e-commerce example,

153
00:09:46,360 --> 00:09:49,680
we can have an orchestrator e-commerce agent

154
00:09:49,680 --> 00:09:54,560
and then we can have an agent specialist in orders management

155
00:09:54,560 --> 00:09:57,680
that can create an order, can cancel an order,

156
00:09:57,680 --> 00:09:59,680
that can get an order.

157
00:09:59,680 --> 00:10:04,800
Another agent can be able to work with a shopping cart,

158
00:10:04,800 --> 00:10:09,160
can add products, remove products to the shopping cart, etc.

159
00:10:09,160 --> 00:10:12,640
Then another agent in terms of product catalog,

160
00:10:12,640 --> 00:10:18,560
of getting a product list or product details.

161
00:10:18,560 --> 00:10:25,120
So all that, what comes up is the, you get a specific agent

162
00:10:25,120 --> 00:10:30,360
that you can, you need to test them, you can optimize

163
00:10:30,360 --> 00:10:34,120
for the task that they are intended for.

164
00:10:34,120 --> 00:10:40,480
And then you can work from a parent agent or an orchestrator agent

165
00:10:40,480 --> 00:10:45,200
and have a multi-agent system, which is much more powerful,

166
00:10:45,200 --> 00:10:49,600
a much more strong than creating one single agent

167
00:10:49,600 --> 00:10:52,680
for a lot of actions and tools.

168
00:10:52,680 --> 00:10:54,560
So that's the idea.

169
00:10:54,560 --> 00:11:01,080
And how do these agents collaborate with one other?

170
00:11:01,080 --> 00:11:04,160
Yeah, so it's about sending messages.

171
00:11:04,160 --> 00:11:08,560
You can, so from the orchestrator,

172
00:11:08,560 --> 00:11:14,080
let's say is the agent that manage all the user input,

173
00:11:14,080 --> 00:11:18,720
then the orchestrator will have, let's say,

174
00:11:18,720 --> 00:11:21,480
in the system instruction or system prompt,

175
00:11:21,480 --> 00:11:27,120
depend on the engine you are using is called instruction or prompt.

176
00:11:27,120 --> 00:11:34,600
You have a description of which sub agents or child agents

177
00:11:34,600 --> 00:11:36,480
or connected agents you have.

178
00:11:36,480 --> 00:11:41,720
And what is the specialization or what is the functions

179
00:11:41,720 --> 00:11:44,760
that each of the agents can do?

180
00:11:44,760 --> 00:11:51,560
So when the user sends a query as a question to the chatbot,

181
00:11:51,560 --> 00:11:57,160
then this orchestrator will decide which of the agents it can be one

182
00:11:57,160 --> 00:12:00,680
or it can be more than one, needs to be called

183
00:12:00,680 --> 00:12:05,080
to create a response for the user query.

184
00:12:05,080 --> 00:12:11,280
And then it will hand off the prompt or the query

185
00:12:11,280 --> 00:12:17,760
to those agents that will perform some actions,

186
00:12:17,760 --> 00:12:21,600
call external tools, whatever it needs to do.

187
00:12:21,600 --> 00:12:27,720
And then it will or generate a response back to the orchestrator

188
00:12:27,720 --> 00:12:33,920
or just perform some actions on a system on databases

189
00:12:33,920 --> 00:12:40,640
or whatever user asks it for.

190
00:12:40,640 --> 00:12:44,320
Then, of course, there is a bunch of ways

191
00:12:44,320 --> 00:12:47,680
that the flow can take.

192
00:12:47,680 --> 00:12:54,160
So if it is, we can let the sub agents to respond

193
00:12:54,160 --> 00:12:56,960
and generate multiple response to the user

194
00:12:56,960 --> 00:13:03,960
or we can extract the orchestrator to gather all responses

195
00:13:03,960 --> 00:13:05,640
and create a unified response.

196
00:13:05,640 --> 00:13:08,880
So that's about different behaviors.

197
00:13:08,880 --> 00:13:13,840
We can manage in the different situations

198
00:13:13,840 --> 00:13:18,360
that our project can have a requirement for.

199
00:13:18,360 --> 00:13:21,040
So we have the orchestration.

200
00:13:21,040 --> 00:13:29,200
So then we have one conductor and a lot of instruments, I think.

201
00:13:29,200 --> 00:13:30,200
Yes.

202
00:13:30,200 --> 00:13:35,160
How does it learn or--

203
00:13:35,160 --> 00:13:41,680
how does the conductor train his employees, his agents?

204
00:13:41,680 --> 00:13:42,720
Yeah.

205
00:13:42,720 --> 00:13:50,320
So the idea is that you can--

206
00:13:50,320 --> 00:13:56,120
even you can have these sub agents collaborating

207
00:13:56,120 --> 00:14:01,080
with different orchestras for different purposes.

208
00:14:01,080 --> 00:14:04,080
So let's say you have the shopping cart.

209
00:14:04,080 --> 00:14:06,720
Or let's say you have the product catalog agent

210
00:14:06,720 --> 00:14:14,520
and you can use coming back to the e-commerce example.

211
00:14:14,520 --> 00:14:18,800
You can use it for external users that

212
00:14:18,800 --> 00:14:22,840
are going to purchase products in your store.

213
00:14:22,840 --> 00:14:25,440
But you can also use that product catalog agent

214
00:14:25,440 --> 00:14:29,840
for your internal employees that will use it

215
00:14:29,840 --> 00:14:35,400
for adding new products to the catalog or improving description.

216
00:14:35,400 --> 00:14:42,760
So that creates like reusable agents

217
00:14:42,760 --> 00:14:48,240
that you can take advantage of in your different projects.

218
00:14:48,240 --> 00:14:50,040
And have we--

219
00:14:50,040 --> 00:14:50,920
I don't know.

220
00:14:50,920 --> 00:14:52,000
Did we--

221
00:14:52,000 --> 00:14:53,360
is this complete autonomous?

222
00:14:53,360 --> 00:14:57,680
Or did we also the human and the look for the kind of--

223
00:14:57,680 --> 00:15:00,440
so that's a good question.

224
00:15:00,440 --> 00:15:04,120
I think that the human needs to be there.

225
00:15:04,120 --> 00:15:09,440
And so there is some actions that you can, of course,

226
00:15:09,440 --> 00:15:14,040
leverage the AI to do it because they are not code actions,

227
00:15:14,040 --> 00:15:21,400
or they are not, let's say, important actions for your company.

228
00:15:21,400 --> 00:15:25,880
But of course, we all know that AI is not deterministic.

229
00:15:25,880 --> 00:15:33,360
So in the case that we need to perform a critical action,

230
00:15:33,360 --> 00:15:37,440
I always recommend that there is an approval process

231
00:15:37,440 --> 00:15:41,360
where a human can approve and can acknowledge

232
00:15:41,360 --> 00:15:44,160
that this action is correct.

233
00:15:44,160 --> 00:15:48,240
Then about the trigger, all these orchestrator

234
00:15:48,240 --> 00:15:50,640
and child agents and connected agents

235
00:15:50,640 --> 00:15:53,640
can be triggered by a human action

236
00:15:53,640 --> 00:15:56,600
or can be triggered by a system event.

237
00:15:56,600 --> 00:16:01,000
So we can have an event-based system

238
00:16:01,000 --> 00:16:07,320
that-- OK, when a product is inserted into the catalog,

239
00:16:07,320 --> 00:16:12,040
there is an event that triggers an agent and this agent

240
00:16:12,040 --> 00:16:17,040
will send a notification for a customer

241
00:16:17,040 --> 00:16:21,080
that was waiting for that product to be added or for that stack

242
00:16:21,080 --> 00:16:25,480
to be replaced.

243
00:16:25,480 --> 00:16:30,680
So the agents can be triggered in an autonomous way.

244
00:16:30,680 --> 00:16:33,480
It's not needed to have a chatbot window

245
00:16:33,480 --> 00:16:35,480
where a user type query and the data

246
00:16:35,480 --> 00:16:38,800
is a response that can be triggered autonomously.

247
00:16:38,800 --> 00:16:42,360
But going to the human in the loop,

248
00:16:42,360 --> 00:16:45,240
I always recommend-- and we always apply

249
00:16:45,240 --> 00:16:49,480
to our customers' projects that the critical operations

250
00:16:49,480 --> 00:16:52,200
must be approved by a human.

251
00:16:52,200 --> 00:16:56,880
Of course, we can make this approval process smoothly.

252
00:16:56,880 --> 00:16:59,920
For instance, if we go to the Microsoft ecosystem,

253
00:16:59,920 --> 00:17:02,960
we have the approval connector in Power Platform, Power

254
00:17:02,960 --> 00:17:08,480
Automate, that we can send an email or a team's notification

255
00:17:08,480 --> 00:17:13,200
to a user with all the information just at a glance.

256
00:17:13,200 --> 00:17:16,160
So the user can just hit approve or deny.

257
00:17:16,160 --> 00:17:17,440
It can put a comment.

258
00:17:17,440 --> 00:17:21,440
And then this will trigger further agents

259
00:17:21,440 --> 00:17:24,640
or further process in the pipeline downstream

260
00:17:24,640 --> 00:17:26,400
to complete the process.

261
00:17:26,400 --> 00:17:28,960
But of course, it's a recommendation

262
00:17:28,960 --> 00:17:34,200
for now that all the critical parts, critical process,

263
00:17:34,200 --> 00:17:35,840
should be approved by a human.

264
00:17:36,480 --> 00:17:47,760
[AUDIO OUT]

265
00:17:47,760 --> 00:17:50,000
Especially in the Microsoft ecosystem,

266
00:17:50,000 --> 00:17:55,640
which tools we have to build in the multi-agent system.

267
00:17:55,640 --> 00:17:58,440
So yeah, there is a lot of them.

268
00:17:58,440 --> 00:18:02,800
And now, currently after Microsoft Build, we have even more.

269
00:18:02,800 --> 00:18:05,720
So let's start by naming in a few.

270
00:18:05,720 --> 00:18:11,880
I'm mostly focused on co-pilot studio and Foundry.

271
00:18:11,880 --> 00:18:16,040
So there, we have different approaches.

272
00:18:16,040 --> 00:18:19,360
We have co-pilot studio agents in which we

273
00:18:19,360 --> 00:18:21,720
can create this--

274
00:18:21,720 --> 00:18:24,840
coming back to the example, this e-commerce orchestrator

275
00:18:24,840 --> 00:18:28,800
with multiple connected agents.

276
00:18:28,800 --> 00:18:32,320
That can be orchestrated by a co-pilot studio agent.

277
00:18:32,320 --> 00:18:37,200
Then we can do the same in Foundry.

278
00:18:37,200 --> 00:18:42,600
So the difference of doing it in co-pilot studio,

279
00:18:42,600 --> 00:18:46,000
doing it in Foundry, it's all about licensing,

280
00:18:46,000 --> 00:18:49,360
all about performance.

281
00:18:49,360 --> 00:18:55,320
So if we have co-pilot studio license in our tenant,

282
00:18:55,320 --> 00:19:00,080
co-pilot studio creates, and then we don't have too much,

283
00:19:00,080 --> 00:19:09,760
like, let's say, too much performance needed.

284
00:19:09,760 --> 00:19:14,840
OK, we don't have too much request or too much users

285
00:19:14,840 --> 00:19:16,160
using the platform.

286
00:19:16,160 --> 00:19:19,520
I will go for co-pilot studio.

287
00:19:19,520 --> 00:19:22,800
You can accelerate the development,

288
00:19:22,800 --> 00:19:29,120
because you have, let's say, more than 1500 connectors,

289
00:19:29,120 --> 00:19:31,160
you can use skills.

290
00:19:31,160 --> 00:19:34,040
Now, with a new orchestrator, you

291
00:19:34,040 --> 00:19:38,880
can use all the power platform ecosystem.

292
00:19:38,880 --> 00:19:43,920
You can have published the agent with a few clicks to WhatsApp,

293
00:19:43,920 --> 00:19:49,240
to Telegram, to Facebook Messenger, Teams,

294
00:19:49,240 --> 00:19:52,280
Microsoft 365 co-pilot, websites,

295
00:19:52,280 --> 00:19:54,400
using the decline channel.

296
00:19:54,400 --> 00:19:58,080
It's very easy to create value quickly.

297
00:19:58,080 --> 00:20:05,040
Then you can go, if you have more performance needs,

298
00:20:05,040 --> 00:20:11,440
or let's say, you need to control the orchestrator face,

299
00:20:11,440 --> 00:20:14,040
you need maybe to create some, I don't know,

300
00:20:14,040 --> 00:20:18,520
group chat patterns or hand-off can be easily done

301
00:20:18,520 --> 00:20:21,800
in co-pilot studio, but let's say more complex patterns,

302
00:20:21,800 --> 00:20:28,000
like, magic, for instance, then you must go to Foundry.

303
00:20:28,000 --> 00:20:33,400
Also there, you have logic apps with a bunch of connectors

304
00:20:33,400 --> 00:20:34,480
too.

305
00:20:34,480 --> 00:20:40,600
You can also control all the orchestration face

306
00:20:40,600 --> 00:20:45,040
using Python code or .NET code or JavaScript,

307
00:20:45,040 --> 00:20:50,880
so you can have more flexibility in what the workflow is

308
00:20:50,880 --> 00:20:52,200
going to be.

309
00:20:52,200 --> 00:20:56,920
And then, for orchestrate using low code,

310
00:20:56,920 --> 00:20:59,360
in co-pilot studio and in Foundry,

311
00:20:59,360 --> 00:21:05,240
we have a new workflow interface, which, where you have a canvas,

312
00:21:05,240 --> 00:21:11,440
and you can connect different agents and put conditional branches

313
00:21:11,440 --> 00:21:14,160
or creating loops between agents.

314
00:21:14,160 --> 00:21:20,640
So you can guide more deterministic process,

315
00:21:20,640 --> 00:21:24,560
yet using AI for the critical parts for, I don't know,

316
00:21:24,560 --> 00:21:28,720
maybe, sumarizing content or document meta-datastraction

317
00:21:28,720 --> 00:21:33,880
or generating some type of text or whatever action,

318
00:21:33,880 --> 00:21:37,760
but you can control all the flow in a deterministic way

319
00:21:37,760 --> 00:21:39,280
using the workflows.

320
00:21:39,280 --> 00:21:41,920
So this is going--

321
00:21:41,920 --> 00:21:44,600
this is in preview now, but we are going

322
00:21:44,600 --> 00:21:52,080
to have it soon in GA and available for our real projects.

323
00:21:52,080 --> 00:22:00,200
And how do connect those actions and workflows play together

324
00:22:00,200 --> 00:22:02,720
in co-pilot studio?

325
00:22:02,720 --> 00:22:03,320
Yeah.

326
00:22:03,320 --> 00:22:07,960
So in co-pilot studio, we currently now in production,

327
00:22:07,960 --> 00:22:10,520
in GA, in general availability, we have what

328
00:22:10,520 --> 00:22:12,320
is called agent flows.

329
00:22:12,320 --> 00:22:14,640
It's similar to Power Automate.

330
00:22:14,640 --> 00:22:18,480
With agent flow, we can trigger the process

331
00:22:18,480 --> 00:22:23,480
using a system event, for instance, when a row is created,

332
00:22:23,480 --> 00:22:27,760
when a row is modified, when some type of email is received,

333
00:22:27,760 --> 00:22:30,480
whatever trigger we can use in Power Automate.

334
00:22:30,480 --> 00:22:36,280
And then we can have a deterministic step-by-step flow.

335
00:22:36,280 --> 00:22:40,240
And we can call agents and AI prompts

336
00:22:40,240 --> 00:22:45,160
and, let's say, different components

337
00:22:45,160 --> 00:22:53,200
who create more deterministic flow and jet using AI

338
00:22:53,200 --> 00:22:55,560
in those points where we need it.

339
00:22:55,560 --> 00:23:00,920
So we used to work like that for our real customer project.

340
00:23:00,920 --> 00:23:04,520
Because in the end, the automations we are creating

341
00:23:04,520 --> 00:23:09,640
in our customer companies is they need to be deterministic.

342
00:23:09,640 --> 00:23:16,240
But AI can be dominate them in specific points.

343
00:23:16,240 --> 00:23:20,480
But if you leverage all the orchestration to AI,

344
00:23:20,480 --> 00:23:25,280
then you convert deterministic need

345
00:23:25,280 --> 00:23:28,120
in an undeterministic process.

346
00:23:28,120 --> 00:23:33,360
And then you are not sure that 100% of the cases,

347
00:23:33,360 --> 00:23:39,440
of the runs of that automation will go in a subset

348
00:23:39,440 --> 00:23:41,320
of response or results.

349
00:23:41,320 --> 00:23:45,520
So that's why we are using these agent flows.

350
00:23:45,520 --> 00:23:51,520
Now, if you go on a preview environment,

351
00:23:51,520 --> 00:23:56,120
that agent flows is called Warflow.

352
00:23:56,120 --> 00:23:58,440
It has a new revamped canvas.

353
00:23:58,440 --> 00:24:02,880
And then you have as a maker, you have more flexibility

354
00:24:02,880 --> 00:24:04,320
to create agents.

355
00:24:07,040 --> 00:24:10,520
There is more simple and more quick way

356
00:24:10,520 --> 00:24:14,080
to create agents that what we have currently

357
00:24:14,080 --> 00:24:16,160
with agent flows.

358
00:24:16,160 --> 00:24:17,960
So yet, we are testing.

359
00:24:17,960 --> 00:24:20,520
We are reporting feedback to Microsoft.

360
00:24:20,520 --> 00:24:28,200
And all together will be much more powerful platform

361
00:24:28,200 --> 00:24:30,680
for AI Warflow.

362
00:24:30,680 --> 00:24:31,880
I hope.

363
00:24:31,880 --> 00:24:34,360
Yeah, that sounds good.

364
00:24:34,360 --> 00:24:41,720
And how companies, I don't know, they have an idea, I think.

365
00:24:41,720 --> 00:24:47,640
How they start to develop a chatbot

366
00:24:47,640 --> 00:24:50,080
or a co-pilot studio?

367
00:24:50,080 --> 00:24:56,880
So we currently have, let's say, an AI advisor program

368
00:24:56,880 --> 00:24:59,960
where we go and we meet the customer.

369
00:24:59,960 --> 00:25:03,280
And we first try to understand

370
00:25:03,280 --> 00:25:04,880
how they work.

371
00:25:04,880 --> 00:25:09,840
What are their, let's say, workflows, internal workflows,

372
00:25:09,840 --> 00:25:12,840
rosters they have in the company?

373
00:25:12,840 --> 00:25:20,240
And we have a team specializes in, let's say,

374
00:25:20,240 --> 00:25:24,880
optimize or create efficiency in those workflows.

375
00:25:24,880 --> 00:25:30,640
So we try to find the critical points

376
00:25:30,640 --> 00:25:35,760
where we can introduce automation.

377
00:25:35,760 --> 00:25:43,960
I need to say that creating automation is not mandatory

378
00:25:43,960 --> 00:25:47,560
to introduce AI in it.

379
00:25:47,560 --> 00:25:54,600
This is one of the most common errors I'm seeing out there.

380
00:25:54,600 --> 00:25:58,200
Because you receive customer requests.

381
00:25:58,200 --> 00:26:00,560
OK, I need, I want to put AI.

382
00:26:00,560 --> 00:26:04,880
In my process, OK, tell me what is your process about?

383
00:26:04,880 --> 00:26:10,080
OK, and then you see that, OK, this is, you need to do this step,

384
00:26:10,080 --> 00:26:11,920
this step, this step, this step.

385
00:26:11,920 --> 00:26:13,080
And that's all.

386
00:26:13,080 --> 00:26:15,120
You don't need AI here.

387
00:26:15,120 --> 00:26:21,880
Even if you put AI, you will create more errors or possible errors

388
00:26:21,880 --> 00:26:26,600
because not always same input, half same output.

389
00:26:26,600 --> 00:26:32,000
So we put, we advise the customer to do the automations.

390
00:26:32,000 --> 00:26:35,520
And we put AI in their specific point

391
00:26:35,520 --> 00:26:40,560
where it creates more value and not more problems.

392
00:26:40,560 --> 00:26:46,320
And so we have this first phase of advisory, of discovery.

393
00:26:46,320 --> 00:26:52,920
And then from that, we create what we call proposals.

394
00:26:52,920 --> 00:26:56,160
And in those proposals, it can be one, can be ten,

395
00:26:56,160 --> 00:26:59,640
can be twenty, whatever number they need.

396
00:26:59,640 --> 00:27:07,320
Then we create some documents that we are using some skills also

397
00:27:07,320 --> 00:27:13,920
to create those documents and to, let's say, make more efficiency

398
00:27:13,920 --> 00:27:17,520
in our internal flows also.

399
00:27:17,520 --> 00:27:20,320
And then we go to the customer again and say, OK,

400
00:27:20,320 --> 00:27:25,040
we propose these ten use cases.

401
00:27:25,040 --> 00:27:31,040
I think that these three ones will take out, I don't know,

402
00:27:31,040 --> 00:27:36,520
maybe 40% of your mechanical work out.

403
00:27:36,520 --> 00:27:38,920
And then we try to start with those that

404
00:27:38,920 --> 00:27:41,520
create more value for the customer.

405
00:27:41,520 --> 00:27:42,720
Awesome.

406
00:27:42,720 --> 00:27:45,000
I think we are both living in the EU.

407
00:27:45,000 --> 00:27:48,240
So we have this EU AI Act.

408
00:27:48,240 --> 00:27:54,920
How can we make these co-pilot studio agents?

409
00:27:54,920 --> 00:27:59,120
Yeah, to be compliant with the EU.

410
00:27:59,120 --> 00:28:03,360
Yeah, this is a tough question, even more in Europe.

411
00:28:03,360 --> 00:28:10,160
So, yeah, of course, we are inside Microsoft ecosystem.

412
00:28:10,160 --> 00:28:15,320
So we have guarantees that we pay the license,

413
00:28:15,320 --> 00:28:20,240
we pay the Azure subscription cost.

414
00:28:20,240 --> 00:28:23,200
And our data is compliant.

415
00:28:23,200 --> 00:28:34,080
For instance, I can tell you in Spain, they passed all the low

416
00:28:34,080 --> 00:28:37,360
requirements in data compliance, in data security

417
00:28:37,360 --> 00:28:39,440
with the Spanish government.

418
00:28:39,440 --> 00:28:42,720
And of course, in the rest of Europe, I think, is the same.

419
00:28:42,720 --> 00:28:48,840
Apart from that, you are guaranteed

420
00:28:48,840 --> 00:28:54,800
that you create an instance of, let's say, GPD model in Azure.

421
00:28:54,800 --> 00:29:02,360
And this instance will be used only by you.

422
00:29:02,360 --> 00:29:06,680
Your data, your problems that you send to the instance

423
00:29:06,680 --> 00:29:13,000
won't be used to train the models back.

424
00:29:13,000 --> 00:29:15,680
This is Microsoft conditions.

425
00:29:15,680 --> 00:29:18,480
So we can sell that to the customer.

426
00:29:18,480 --> 00:29:24,600
Then there is a new player, which is anthropic.

427
00:29:24,600 --> 00:29:30,640
And of course, we know all the power of the cloud models.

428
00:29:30,640 --> 00:29:32,520
We like all the cloud models.

429
00:29:32,520 --> 00:29:36,480
But we also know that cloud models for the moment,

430
00:29:36,480 --> 00:29:42,320
and I pay this to be changed, are running in US data center.

431
00:29:42,320 --> 00:29:45,760
We also know that all the communication is encrypted.

432
00:29:45,760 --> 00:29:50,080
Only the inference is done in US.

433
00:29:50,080 --> 00:29:55,280
But of course, depending on the customer, this can be a problem.

434
00:29:55,280 --> 00:29:59,360
So for those customers, what we do is, OK, we

435
00:29:59,360 --> 00:30:01,320
is playing the situation.

436
00:30:01,320 --> 00:30:03,120
We are crystal clear with them.

437
00:30:03,120 --> 00:30:09,160
And then we just swap to GPD models.

438
00:30:09,160 --> 00:30:14,120
Now there is a new player, Mistral, that

439
00:30:14,120 --> 00:30:17,000
is coming into the Microsoft tenants.

440
00:30:17,000 --> 00:30:20,600
You can enable, and those models are running in Europe.

441
00:30:20,600 --> 00:30:25,600
So we have now two processors for choose--

442
00:30:25,600 --> 00:30:29,320
OpenAI, Mistral in Europe, and Cloud.

443
00:30:29,320 --> 00:30:34,880
We know that for a moment is in US.

444
00:30:34,880 --> 00:30:38,640
But yeah, that's all about--

445
00:30:38,640 --> 00:30:41,880
we know-- we work with Microsoft Technologies,

446
00:30:41,880 --> 00:30:44,240
and we know that we are compliant.

447
00:30:44,240 --> 00:30:48,320
We need to pay the price of Azure subscription of licensing.

448
00:30:48,320 --> 00:30:52,440
But that guarantees all that our data is secure,

449
00:30:52,440 --> 00:30:56,960
and won't be used for training, again, back to models.

450
00:30:56,960 --> 00:31:06,040
Like you are using Anthropic or OpenAI,

451
00:31:06,040 --> 00:31:08,920
this can happen, of course.

452
00:31:08,920 --> 00:31:10,360
Yeah.

453
00:31:10,360 --> 00:31:12,240
We have one side.

454
00:31:12,240 --> 00:31:14,920
We have this co-pilot studio, and the other side,

455
00:31:14,920 --> 00:31:16,840
we have Microsoft Foundry.

456
00:31:16,840 --> 00:31:20,080
What are the difference between both?

457
00:31:20,080 --> 00:31:25,560
And when I choose co-pilot, or when I choose Microsoft Foundry?

458
00:31:25,560 --> 00:31:30,200
Yeah, so the difference are closing the gap.

459
00:31:30,200 --> 00:31:31,840
So Microsoft is closing the gap.

460
00:31:31,840 --> 00:31:36,120
I remember when first co-pilot studio was launched,

461
00:31:36,120 --> 00:31:39,280
and there is even no Foundry.

462
00:31:39,280 --> 00:31:43,080
They were called Azure OpenAI, that, of course,

463
00:31:43,080 --> 00:31:45,960
there were big, big difference.

464
00:31:45,960 --> 00:31:50,440
Co-pilot studio was more low-code approach.

465
00:31:50,440 --> 00:31:52,840
And then if you need more flexibility,

466
00:31:52,840 --> 00:31:56,480
you need to go Foundry, or let's say, Azure OpenAI,

467
00:31:56,480 --> 00:31:58,000
in backing days.

468
00:31:58,000 --> 00:32:00,600
Now, they are closing the gap.

469
00:32:00,600 --> 00:32:05,320
Let's say you can go Foundry with a low-code approach,

470
00:32:05,320 --> 00:32:08,720
because you even, if you don't need so much flexibility

471
00:32:08,720 --> 00:32:11,160
in the orchestration part, you can go Foundry,

472
00:32:11,160 --> 00:32:15,480
you can create multi-agent systems using the low-code approach

473
00:32:15,480 --> 00:32:16,320
in Foundry.

474
00:32:16,320 --> 00:32:24,560
So if we go low-code, we need to think about what our customer,

475
00:32:24,560 --> 00:32:27,840
or our company, if we are developing a custom product

476
00:32:27,840 --> 00:32:32,680
for our company, which type of licensing,

477
00:32:32,680 --> 00:32:36,080
or which type of, where is the data located?

478
00:32:36,080 --> 00:32:38,880
Are we using Power Platform?

479
00:32:38,880 --> 00:32:41,600
We have all data in Databrers.

480
00:32:41,600 --> 00:32:45,960
We have Power Platform license, co-pilot studio credits.

481
00:32:45,960 --> 00:32:48,920
Then of course, we go co-pilot studio.

482
00:32:48,920 --> 00:32:51,640
Are we using Azure?

483
00:32:51,640 --> 00:32:57,800
We have external information that we need to connect to.

484
00:32:57,800 --> 00:33:00,000
We don't have any Power Platform.

485
00:33:00,000 --> 00:33:02,120
We don't use Databrers, of course, then.

486
00:33:02,120 --> 00:33:04,280
We can go Foundry.

487
00:33:04,280 --> 00:33:06,120
Apart from that, this is low-code approach.

488
00:33:06,120 --> 00:33:11,080
If we go Pro-code approach, then we have agent framework.

489
00:33:11,080 --> 00:33:15,320
And agent framework will work much, much better with Foundry.

490
00:33:15,320 --> 00:33:22,720
So of course, if we need to create a much more custom orchestration,

491
00:33:22,720 --> 00:33:27,920
much more custom agents, maybe make different calls,

492
00:33:27,920 --> 00:33:32,080
specific workflows, then we go Pro-code.

493
00:33:32,080 --> 00:33:37,080
And then we use Foundry models, of course.

494
00:33:37,080 --> 00:33:42,080
So that would be my top decision.

495
00:33:42,080 --> 00:33:48,080
Then of course, specific parts or specific requirements of the customer

496
00:33:48,080 --> 00:33:53,080
can make you take one decision or another.

497
00:33:53,080 --> 00:34:00,080
If you go co-pilot studio, then you think about also the,

498
00:34:00,080 --> 00:34:08,080
let's say how many users will be using the product.

499
00:34:08,080 --> 00:34:16,080
And then you can also estimate how much credits you need to buy.

500
00:34:16,080 --> 00:34:22,080
If you plan to buy a credit package, or if you plan to go pay as you go,

501
00:34:22,080 --> 00:34:29,080
you can try to estimate if your number of requests

502
00:34:29,080 --> 00:34:34,080
or number of users using the platform will be more expensive

503
00:34:34,080 --> 00:34:39,080
than going in a Foundry approach.

504
00:34:39,080 --> 00:34:43,080
But yeah, they are closing the gap.

505
00:34:43,080 --> 00:34:48,080
Yesterday I was working in a co-pilot studio agent that connects directly

506
00:34:48,080 --> 00:34:55,080
to Foundry IQ, knowledge base using the MCP connector.

507
00:34:55,080 --> 00:35:03,080
So before last week for doing that, for connecting a co-pilot studio agent

508
00:35:03,080 --> 00:35:12,080
to a Foundry IQ knowledge base, you need to go through a Foundry agent.

509
00:35:12,080 --> 00:35:16,080
So there you lose flexibility in the response.

510
00:35:16,080 --> 00:35:23,080
But now you can connect using a MCP connector directly,

511
00:35:23,080 --> 00:35:27,080
the co-pilot studio agent to a Foundry IQ.

512
00:35:27,080 --> 00:35:30,080
So it's even more close the gap.

513
00:35:30,080 --> 00:35:37,080
Now, if you have Foundry IQ, you don't need to go and create a Foundry agent.

514
00:35:37,080 --> 00:35:39,080
You can go and create a co-pilot studio.

515
00:35:39,080 --> 00:35:43,080
So what Microsoft is trying to do is closing the gap

516
00:35:43,080 --> 00:35:49,080
and just you choose base on your, let's say,

517
00:35:49,080 --> 00:35:54,080
on your current resources or your current technology

518
00:35:54,080 --> 00:36:01,080
that you're using in the rest of your project and not because one platform

519
00:36:01,080 --> 00:36:06,080
have different, let's say, functionalities or features

520
00:36:06,080 --> 00:36:09,080
that are missing in the another platform.

521
00:36:09,080 --> 00:36:11,080
That's how I see it.

522
00:36:11,080 --> 00:36:14,080
Yeah, you have a Microsoft agent framework.

523
00:36:14,080 --> 00:36:18,080
What is the Microsoft agent framework?

524
00:36:18,080 --> 00:36:21,080
And how can we explain it to people

525
00:36:21,080 --> 00:36:24,080
that are not familiar with this concept?

526
00:36:24,080 --> 00:36:33,080
Yeah, so agent framework comes from two different teams inside Microsoft

527
00:36:33,080 --> 00:36:38,080
that were creating in the past was called semantic kernel.

528
00:36:38,080 --> 00:36:46,080
And then those teams join it to create a unified framework

529
00:36:46,080 --> 00:36:51,080
because, of course, the community or the customers

530
00:36:51,080 --> 00:36:57,080
were creating projects using two different tools from Microsoft

531
00:36:57,080 --> 00:37:01,080
that were intended to be for the same purpose.

532
00:37:01,080 --> 00:37:03,080
So they create the agent framework.

533
00:37:03,080 --> 00:37:09,080
And this is all about creating agents is a framework

534
00:37:09,080 --> 00:37:14,080
that you can use from different programming languages.

535
00:37:14,080 --> 00:37:21,080
I think they currently have SDKs for JavaScript, for C# and Python.

536
00:37:21,080 --> 00:37:25,080
Maybe there's another that I don't know.

537
00:37:25,080 --> 00:37:31,080
Then you can create agents using Pro Code.

538
00:37:31,080 --> 00:37:35,080
So you can create all the features that an agent has.

539
00:37:35,080 --> 00:37:40,080
You can create a system from you can add tools or functions or skills

540
00:37:40,080 --> 00:37:42,080
or whatever you have.

541
00:37:42,080 --> 00:37:49,080
And then you can create also workflows or orchestration patterns.

542
00:37:49,080 --> 00:37:56,080
You can add those agents to the workflows and make them work together

543
00:37:56,080 --> 00:37:59,080
for specific purpose.

544
00:37:59,080 --> 00:38:04,080
And then you can host that agents inside Foundry, for instance,

545
00:38:04,080 --> 00:38:09,080
or you can publish to Azure Container Up in Azure.

546
00:38:09,080 --> 00:38:16,080
And then you have a API or a user interface where you can interact with agents.

547
00:38:16,080 --> 00:38:22,080
So that gives you, I think, is the more flexible way of creating agents.

548
00:38:22,080 --> 00:38:28,080
You never hit a wall there because in Copa I was studio and Foundry low code approach.

549
00:38:28,080 --> 00:38:30,080
Just sometimes hit a wall.

550
00:38:30,080 --> 00:38:32,080
I'm in my first project.

551
00:38:32,080 --> 00:38:37,080
It happens to me that, okay, we have these requirements.

552
00:38:37,080 --> 00:38:42,080
And then the customer requests another feature that was not in the primary,

553
00:38:42,080 --> 00:38:44,080
let's say, requirement list.

554
00:38:44,080 --> 00:38:51,080
And then we say, oh my God, we cannot do that in Copa I was studio or we cannot do that in Foundry low code.

555
00:38:51,080 --> 00:38:53,080
We need to go Pro Code.

556
00:38:53,080 --> 00:38:57,080
So now, of course, we know better the platform.

557
00:38:57,080 --> 00:39:02,080
We know better the limits and the walls where they are, what they are.

558
00:39:02,080 --> 00:39:08,080
So we can guide or we can decide better from the beginning.

559
00:39:08,080 --> 00:39:19,080
But of course, if you have Dabs or you have some plan to go extend your project in a more customized way,

560
00:39:19,080 --> 00:39:23,080
then you can go from the beginning to agent framework.

561
00:39:23,080 --> 00:39:30,080
You can build exactly the same that you build on Copa I studio or Foundry low code approach.

562
00:39:30,080 --> 00:39:35,080
Of course, you will take a little bit more effort, a little bit more resources,

563
00:39:35,080 --> 00:39:46,080
but then you have a strong base for growing up your agents ecosystem without any limit.

564
00:39:46,080 --> 00:39:47,080
Awesome.

565
00:39:47,080 --> 00:39:53,080
Yeah, I think there are a lot of possibilities of Microsoft there, there.

566
00:39:53,080 --> 00:39:58,080
But we have also a topic called fine ops.

567
00:39:58,080 --> 00:40:02,080
How we have the control over the coast.

568
00:40:02,080 --> 00:40:04,080
Can you a little bit explain this?

569
00:40:04,080 --> 00:40:05,080
Yeah.

570
00:40:05,080 --> 00:40:09,080
This is very difficult to control.

571
00:40:09,080 --> 00:40:13,080
I can put an example on Microsoft.

572
00:40:13,080 --> 00:40:15,080
So we will be forcing in every project.

573
00:40:15,080 --> 00:40:19,080
We have that question from the customer.

574
00:40:19,080 --> 00:40:23,080
What will be the cause of this when we use it?

575
00:40:23,080 --> 00:40:27,080
Okay, and we say it depends.

576
00:40:27,080 --> 00:40:31,080
Hey, you tell me you have 100 users.

577
00:40:31,080 --> 00:40:36,080
Let's say, but how much times they will use the platform?

578
00:40:36,080 --> 00:40:40,080
How much queries they have?

579
00:40:40,080 --> 00:40:43,080
Which type of queries they will launch?

580
00:40:43,080 --> 00:40:51,080
Let's say you have a agent, let's put one agent with five tools.

581
00:40:51,080 --> 00:40:56,080
Let's say that we have 10 type of queries.

582
00:40:56,080 --> 00:41:00,080
And this type of queries will launch one tool.

583
00:41:00,080 --> 00:41:03,080
Another type of query will launch three tools.

584
00:41:03,080 --> 00:41:06,080
Another type of query will launch five tools.

585
00:41:06,080 --> 00:41:16,080
So if you make 10 queries that launch five tools each query,

586
00:41:16,080 --> 00:41:20,080
then you have 50 tools launch.

587
00:41:20,080 --> 00:41:28,080
This will cost more than if the same user launched 10 queries that will launch one tool.

588
00:41:28,080 --> 00:41:31,080
This is a total of 10 tools.

589
00:41:31,080 --> 00:41:33,080
And this will cost less.

590
00:41:33,080 --> 00:41:41,080
So it's very difficult to calculate that a Microsoft have a launch a copilot to your estimator.

591
00:41:41,080 --> 00:41:56,080
But there is a big sign in the top that says, okay, we don't make responsible to what this estimator will tell you because the cost of course depends on the usage.

592
00:41:56,080 --> 00:42:02,080
So yeah, you can make maybe an average.

593
00:42:02,080 --> 00:42:11,080
I think that for me, what I recommend to the customer is, okay, I can give you, let's say, an average of,

594
00:42:11,080 --> 00:42:22,080
we and the UAT face or testing face, we can make a bunch of tests and then we can see the cost, what it is.

595
00:42:22,080 --> 00:42:39,080
We can also have pre-built estimations because because we have some tables in the documentation that say, okay, one message that calls the order, the generative,

596
00:42:39,080 --> 00:42:47,080
or the structure and then calls three tools, it will cost, I don't know, 10 credits or 50 credits.

597
00:42:47,080 --> 00:42:59,080
Okay, we can make some estimation, but these are not going to any place because in the end, the usage, the real usage is what we miss or what it miss or the cost.

598
00:42:59,080 --> 00:43:09,080
So in the testing face, we have ways to get the total consumption credits, we can send that to the customer.

599
00:43:09,080 --> 00:43:24,080
And then when we go live, we miss or the first month or let's say we agree with a period of miss or meant with the customer and then we adapt the resources or the,

600
00:43:24,080 --> 00:43:42,080
if we are in Copa de Studio, we can see what number of credits we consume in a month or if we need to plug as a subscription to go pay as you go, in case the credits pack, which is 35,000 will be,

601
00:43:42,080 --> 00:43:46,080
will not reach one month.

602
00:43:46,080 --> 00:44:01,080
And then on the on the file, you have a token based consumption, you can miss or let's say in the play round of the agent, you can miss or how many tokens are taking up your agent.

603
00:44:01,080 --> 00:44:15,080
And then of course, you can have estimation of, okay, you send this query 100 times, you will consume roughly this number of tokens.

604
00:44:15,080 --> 00:44:23,080
But so there is no silver wallet for this is very difficult.

605
00:44:23,080 --> 00:44:31,080
We need to educate the customer to not waste tokens.

606
00:44:31,080 --> 00:44:44,080
We see this month, they did have copilot price increasing or changing from rate to user base.

607
00:44:44,080 --> 00:44:48,080
And this is what we have.

608
00:44:48,080 --> 00:44:59,080
Until now, we are paying 10% of what AI is costing to the companies that provide it.

609
00:44:59,080 --> 00:45:08,080
And now we are, we need to learn to save cost to optimize the usage.

610
00:45:08,080 --> 00:45:13,080
We as a builders and we as a users.

611
00:45:13,080 --> 00:45:17,080
So, I know, I put you a, a simple example.

612
00:45:17,080 --> 00:45:35,080
We used to, when we developed a front end project, we have this MPN run comment for starting the back or for compiling the assets of a project.

613
00:45:35,080 --> 00:45:44,080
And the one line comment that we put in the terminal, all the, all in all our life, we have put that line in the terminal.

614
00:45:44,080 --> 00:45:50,080
But now people is asking the agent to run that.

615
00:45:50,080 --> 00:45:55,080
And this is, I don't know, 200 tokens or 1000 tokens that is running away.

616
00:45:55,080 --> 00:46:04,080
Why we don't put ourself the comment in the terminal like we have been done the past 10 years.

617
00:46:04,080 --> 00:46:10,080
So, it's like we are getting more lazy with AI.

618
00:46:10,080 --> 00:46:22,080
But I think this will change now because our pocket will, will be requesting us to save tokens and, yeah, an Indian save the planet.

619
00:46:22,080 --> 00:46:37,080
And, and have we, yeah, I think, especially when we think about multi agent, there are new risk, especially I think governance or compliance or cybersecurity.

620
00:46:37,080 --> 00:46:40,080
We have to look about.

621
00:46:40,080 --> 00:46:51,080
This is a very current topic. We are, I'm just there. I was working with a colleague in connecting with agent 365.

622
00:46:51,080 --> 00:47:08,080
We, we can have what Microsoft call a control plane, which is, okay, we can have I agents in co-pilot studio. We can have Microsoft agent framework agents deployed in Azure container apps.

623
00:47:08,080 --> 00:47:24,080
We can have found the agents. But I go further, we can have Amazon bedrock agents. And we can connect all these telemetry together in agents 365 control plane.

624
00:47:24,080 --> 00:47:38,080
We can have there, we can block the agents, we can have all the tools that they are using, how much traffic they have, we can have all the telemetry in one place.

625
00:47:38,080 --> 00:47:57,080
So this will be Microsoft before is working. Just today we were testing this, this feature that my colleague will, we will, we will showing that in the European PowerPoint conference in the end of the month.

626
00:47:57,080 --> 00:48:16,080
And yeah, it's, it's very powerful. It's not yet fully working. Let's say some agents are not reporting all the telemetry, but we are all working together in reporting and improving the tool.

627
00:48:16,080 --> 00:48:29,080
So this will be the centralized Microsoft admin center tool for govern all the agents that we have in the big Microsoft ecosystem.

628
00:48:29,080 --> 00:48:34,080
And yeah, there is a, it's a good, good topic.

629
00:48:34,080 --> 00:48:51,080
And of course, if is one of the things that we need to care about, even, even more, if we have different teams in the company that can build agents by themselves.

630
00:48:51,080 --> 00:49:14,080
We have learned from the power platform, power, as power automate some years ago, what we call, we had the center of excellence that help us to govern all the apps and all the flows that were created inside our tenant inside our environments.

631
00:49:14,080 --> 00:49:37,080
Now, with all that lesson that we have learned, we create or Microsoft create this agent 365 control plane that is helping right now to govern the agents and is improving by week every week with new feature with new telemetries and, yeah, I think is very powerful.

632
00:49:37,080 --> 00:49:40,080
This is the way for control and govern the events.

633
00:49:40,080 --> 00:49:45,080
So then, yeah, let's jump into the quick fire round.

634
00:49:45,080 --> 00:49:53,080
I asked, Jen, then you give a short answer or what come first in your mind, cool, but it's do or as a fall, agree.

635
00:49:53,080 --> 00:49:54,080
Co-piles to you.

636
00:49:54,080 --> 00:50:04,080
Anyway, I will open the eye.

637
00:50:04,080 --> 00:50:09,080
Most underrated Microsoft technology.

638
00:50:09,080 --> 00:50:12,080
Power pages.

639
00:50:12,080 --> 00:50:21,080
When you have the chance of Microsoft's release tomorrow, future or to what should be.

640
00:50:21,080 --> 00:50:35,080
Let's say.

641
00:50:35,080 --> 00:50:41,080
Let's say.

642
00:50:41,080 --> 00:50:50,080
Web, some web technology that help to include AI in their websites.

643
00:50:50,080 --> 00:51:00,080
When someone comes to you to the Canadian island, what is the favorite meal they shall take?

644
00:51:00,080 --> 00:51:06,080
Oh, I'm potatoes.

645
00:51:06,080 --> 00:51:13,080
Yeah, have you one favorite productivity tip?

646
00:51:13,080 --> 00:51:19,080
Productivity.

647
00:51:19,080 --> 00:51:21,080
Yeah.

648
00:51:21,080 --> 00:51:23,080
Yeah.

649
00:51:23,080 --> 00:51:42,080
So I think that using AI or let's say, I put a concrete example, copilot, co-work for organizing your day, your daily task and creating skills that helps you to automate mechanical tasks.

650
00:51:42,080 --> 00:51:52,080
When you meet your 20 years, stinger yourself, what advisable you give him?

651
00:51:52,080 --> 00:52:10,080
Yeah, so currently now it's learned to use AI for your task or your daily work, whatever it be, you need to get more efficiency.

652
00:52:10,080 --> 00:52:16,080
For improvement over the years.

653
00:52:16,080 --> 00:52:21,080
And is there any learning source, book or conference?

654
00:52:21,080 --> 00:52:27,080
You say the lists where every in textual visits read or yeah.

655
00:52:27,080 --> 00:52:28,080
Yeah.

656
00:52:28,080 --> 00:52:34,080
So I will recommend to learn.

657
00:52:34,080 --> 00:52:45,080
So I think now the books are a bit outdated in the sense that technology is evolving too fast.

658
00:52:45,080 --> 00:52:58,080
So I usually go and find some good YouTube channels that will be creating content in a weekly or monthly base.

659
00:52:58,080 --> 00:53:07,080
So for me, a compiler studio, there is a good channel that is from Dwayne Robinson, Microsoft Guy.

660
00:53:07,080 --> 00:53:22,080
But of course, depending on your skills, on your needs, you can find, but let's say a good YouTube channel with that create content for your area.

661
00:53:22,080 --> 00:53:28,080
And that is this content is updating frequently.

662
00:53:28,080 --> 00:53:39,080
Okay, then my closing question is if listeners remember one thing from this conversation today about AI genes and intelligent automation, what should it be?

663
00:53:39,080 --> 00:53:40,080
Yeah.

664
00:53:40,080 --> 00:53:59,080
So for me, one closing phrase that I will take is let's imagine you are in a beach and you have a surfing board in your hand and you watch the sea and is a big wave coming to you.

665
00:53:59,080 --> 00:54:03,080
This big wave is AI and then you have two options.

666
00:54:03,080 --> 00:54:10,080
You stay and the wave we go over you or you take the surfing board and surf the wave.

667
00:54:10,080 --> 00:54:22,080
So I recommend you do the second surf the wave, take AI on your hand and try to be more efficient in your daily work in your life.

668
00:54:22,080 --> 00:54:36,080
So yeah, then I say thank you David. This was really great conversation especially about co-party studio Microsoft, founder of the Microsoft agent framework and multi agent AI systems.

669
00:54:36,080 --> 00:54:43,080
Yeah, I hope you have enjoyed this discussion and I say yeah, thank you so much for joining me and have a nice day.

670
00:54:43,080 --> 00:54:44,080
Bye.

671
00:54:44,080 --> 00:54:45,080
Thank you.

672
00:54:45,080 --> 00:54:46,080
Thank you.

673
00:54:46,080 --> 00:55:14,220
[知 away from mic]

Mirko Peters

Founder of m365.fm, m365.show and m365con.net

Mirko Peters is a Microsoft 365 expert, content creator, and founder of m365.fm, a platform dedicated to sharing practical insights on modern workplace technologies. His work focuses on Microsoft 365 governance, security, collaboration, and real-world implementation strategies.

Through his podcast and written content, Mirko provides hands-on guidance for IT professionals, architects, and business leaders navigating the complexities of Microsoft 365. He is known for translating complex topics into clear, actionable advice, often highlighting common mistakes and overlooked risks in real-world environments.

With a strong emphasis on community contribution and knowledge sharing, Mirko is actively building a platform that connects experts, shares experiences, and helps organizations get the most out of their Microsoft 365 investments.

David Lorenzo Lopez

MVP BizzApps - Automation & AI Technical Architect

Father of 2 princess and husband of a queen.
Since 2010 I've helped organizations turn business challenges into working solutions, across enterprise development and the cloud. Last 5 years I've had a sharper focus on intelligent automation and AI-driven ways of working. My work centers on helping teams adopt new technology in a way that actually moves the needle: faster processes, better decisions, and real outcomes.
Beyond the day job, one of my greatest passions is sharing knowledge with the community. I speak at events around the world, learn constantly from other professionals, and take particular pride in mentoring those taking their first steps in this field, guiding them along the way and watching them grow into outstanding professionals.
When I'm not working, you'll find me enjoying football, playing guitar, listening to music, and spending time with family and friends.
You can reach me at [email protected]

david