The AB-620 training is for people who have already built an agent in Microsoft Copilot Studio and are stuck on the next step: connecting it to the systems the business actually runs on. An agent that answers from a handful of documents is no longer enough. It has to query an ERP, trigger an action in a business tool, hand a task to another agent and stay governed. This course teaches you to design those integrated solutions, from the architecture choice through to deployment.
The official Microsoft curriculum, rebuilt in October 2026, has 18 modules. You discover Copilot Studio agent harnesses, including the GitHub Copilot harness, and how to keep their costs under control. You create instruction-driven agents, add reusable skills, ground answers in enterprise knowledge and Microsoft IQ, then extend the agent with tools, MCP servers and human-in-the-loop workflows. The final part covers multi-agent solutions, governance, responsible AI, evaluation, monitoring and ALM. You leave prepared for the AB-620 exam.
Module 1: Explore Microsoft Copilot Studio and agent harnesses
Module 2: Manage usage and cost for agents in Microsoft Copilot Studio
Module 3: Create instruction-driven agents in Microsoft Copilot Studio
Module 4: Design reusable behavior with skills in Microsoft Copilot Studio
Module 5: Design integration strategies for agents in Microsoft Copilot Studio (GitHub Copilot)
Module 6: Automate agent tasks with workflows in Microsoft Copilot Studio
Module 7: Add human-in-the-loop steps to workflows in Microsoft Copilot Studio
Module 8: Ground agents in enterprise knowledge in Microsoft Copilot Studio (GitHub Copilot)
Module 9: Extend agents with tools in Microsoft Copilot Studio (GitHub Copilot)
Module 10: Connect agents to MCP servers in Microsoft Copilot Studio (GitHub Copilot)
Module 11: Connect agents to Microsoft IQ in Microsoft Copilot Studio (GitHub Copilot)
Module 12: Design multi-agent solutions in Microsoft Copilot Studio
Module 13: Build connected agent solutions in Microsoft Copilot Studio (GitHub Copilot)
Module 14: Design a governance and security strategy for agents in Microsoft Copilot Studio
Module 15: Design a responsible AI strategy for agents in Microsoft Copilot Studio
Module 16: Evaluate agents in Microsoft Copilot Studio (GitHub Copilot)
Module 17: Monitor and improve agents in Microsoft Copilot Studio (GitHub Copilot)
Module 18: Implement application lifecycle management for agents in Microsoft Copilot Studio with Power Platform
The first wave of Copilot Studio projects produced plenty of question-and-answer agents: a document library, a few instructions, a Teams publication. They help, but they stop at the moment the user expects an action. Checking an order status, opening a ticket, booking a resource or comparing figures from two systems all require the agent to call something other than its documents. That is the boundary the AB-620 training crosses.
The course starts from a simple idea: an enterprise agent is first an integration problem. You have to pick the right connection method, handle the identity of the user or service acting, limit what the agent is allowed to do and check that it does it correctly. The curriculum spends a full module on these integration choices before going into connectors, REST APIs and MCP in detail.
The curriculum opens with a concept that is new in Copilot Studio: the agent harness, the runtime that orchestrates the model, its knowledge and its tools. You compare the available harnesses, including the GitHub Copilot harness, and their billing models, before building instruction-driven agents enriched with reusable skills. Grounding relies on enterprise knowledge sources and on Microsoft IQ: Foundry IQ for knowledge bases, Fabric IQ for business data and Work IQ for Microsoft 365 context.
Two modules then cover composing several agents. You prepare an agent so others can connect to it, write the routing description that guides delegation and call an agent from a workflow.
The trainer insists on the question that comes before the technology: do you really need several agents? A multi-agent solution costs more to test and monitor. The design module gives you the criteria to decide, and the coordination patterns to use when the answer is yes.
The Microsoft path around agents has three levels:
If you build Power Platform extensions in code rather than agents, the Extend Microsoft Power Platform Solutions with code and AI (AB-400) training is a better fit.
The AB-620 exam leads to the Microsoft Certified: AI Agent Builder Associate certification. It assesses three areas: planning and configuring agent solutions, integrating and extending them, then testing and lifecycle management. That last part, often neglected in projects, weighs a quarter of the exam: test sets, evaluation methods, solutions, environment variables and Power Platform pipelines. The exam is booked with Microsoft and remains optional.
Bring a real case: a process where an agent should act in a system you know, with its access constraints. The exercises run in the official Microsoft labs, but every integration module raises the same question you will face at home: who authenticates, with which rights, and what happens when the call fails? Leaving with those answers for your own case is worth as much as the exercises themselves. The labs stay available for 180 days, so you can replay an integration before reproducing it in your own environment.
Do I need PL-7008 before AB-620?
It is not mandatory, but AB-620 assumes you can already build an agent with its instructions, knowledge and tools. PL-7008 covers exactly that foundation in one day.
Do I need to know how to code?
The course stays in Copilot Studio and does not ask you to write an application. You do need to be comfortable with REST APIs, OpenAPI specifications and authentication concepts, because a large part of the curriculum is about connecting the agent to external services.
How is it different from AI-500?
AI-500 is for developers who build multi-agent systems in code on Azure and Microsoft Foundry. AB-620 builds the same kinds of solutions from Copilot Studio, relying on its connectors, flows and governance.
Does the course cover MCP?
Yes, a full module covers connecting an agent to an existing MCP server, authenticating that connection and managing the MCP tools exposed to the agent.
How much Power Platform knowledge do I need?
Copilot Studio is built on Power Platform: agents live in environments, rely on Dataverse and are deployed through solutions. Being comfortable with Power Fx, environments and connectors will save you time, especially in the part dedicated to testing and lifecycle management.
Which licences do I need to reproduce the exercises at work?
The exercises use the official Microsoft labs, so no licence is needed during the training. For your own environment, the trainer reviews what Copilot Studio, premium connectors and Microsoft IQ services require.
No Events
Monday to Friday
8:30 AM to 6:00 PM
Tel. 058 307 73 00
ITTA
Route des jeunes 35
1227 Carouge, Suisse
Monday to Friday, from 8:30 am to 06:00 pm.
