AB-620 on Microsoft Learn: learning path, labs and practice assessment (what exists and what does not)

The AB-620 study guide is the reference document: candidate profile, the three domains with their weights (plan and configure 30–35%, integrate and extend 40–45%, test and manage 20–25%) and the bullets of every sub-skill, from "plan identity strategy" to "create a multi-agent solution by using A2A protocol". It has been in force since 9 April 2026 and it is the source the questions come from; read it whole and use it as a checklist.

The exam sandbox shows the real interface and question types (single, multiple, drag, order) with generic content. And the guide links the documentation that matters: Copilot Studio, Microsoft 365 Copilot, Power Platform and Microsoft Foundry, plus the community forums and Microsoft Learn Shows. The AB-620 outline explained summarizes that guide with an hour split.

Microsoft has published course AB-620T00-A, "Design and build integrated AI agent solutions in Copilot Studio": intermediate level, three days, in English, with the instructor-led version announced for 30 September 2026. It is built on three free learning paths: designing agent conversations, responses and topics; designing and building multi-agent solutions; and integrating agents with enterprise systems. It is the best study order available and mostly covers domains 1 and 2. What still does not exist is the free practice assessment: the page says it is not currently available and that they are usually published within eight weeks of the exam leaving beta.

Two more official resources help: the Copilot Studio labs repository Microsoft maintains on GitHub (microsoft/mcs-labs), built for workshops but useful for practice, and the Applied Skills credential "Build an agent in Microsoft Copilot Studio", which you earn through an assessed lab. And the Copilot Studio documentation covers every bullet of the outline: agent flows and human approvals, topics, adaptive cards, connectors and knowledge sources, MCP tools and resources, computer use, connecting Foundry and Fabric agents, A2A, evaluation with test sets, solutions and pipelines. Read it by sub-skill, not by product.

You need a Power Platform developer environment (free) with Copilot Studio, a Microsoft Foundry project in a trial Azure subscription, and a simple API of your own (an Azure Function or a test service) for integrations. Domain 1: create an agent with instructions and a knowledge source, add an agent flow with a human approval and an adaptive card that collects data; configure Entra authentication and publish it to Teams.

Domain 2: connect the agent to Azure AI Search, add a custom connector over your API, expose an MCP server with one tool and one resource, integrate a Foundry agent and try an A2A connection. Domain 3: create a test set with the three evaluation methods, enable Application Insights, put the agent in a solution with environment variables and deploy it to another environment with a pipeline. With that you have touched everything the exam asks.

The order that works best is to read one sub-skill of the guide, do the matching lab and then answer questions from that domain to check whether you can tell the components apart in scenarios. Repeating the cycle across the three domains takes four to eight weeks depending on your profile; the plan by profile tells you how many.

Without an official practice assessment, question practice is the only thing that tells you whether you are ready. The AB-620 practice test has tests by domain with explanations and a 120-minute mock exam with the official weights; start for free, no sign-up, and book the exam when your results stay stable on mock exams with unseen questions.