AB-620 Domains Explained With Weightings

The short version

The study guide has three domains and nine skill areas. The middle domain, integrating and extending agents, carries 40-45% of the exam, so it deserves the most of your time. The traps are rarely exotic. They come from two similar features or from a setting that behaves differently from what you assumed. Weightings are Microsoft’s, from the study guide on Microsoft Learn (last updated October 16, 2026 according to its page). The “traps” and “practice” notes below are our own advice, not Microsoft statements.

Domain	Weighting
Plan and configure agent solutions	30-35%
Integrate and extend agents in Copilot Studio	40-45%
Test and manage agents	20-25%

Domain 1: Plan and configure agent solutions (30-35%)

Plan an agent solution

Microsoft lists integration with enterprise systems, identity strategy, channels and deployment, responsible AI, security and governance, reusable components, and designing for internal or external audiences. In practice you choose how users sign in, where the agent is published, and what an administrator can block.

Typical trap: data policies. Connectors sit in groups, data cannot be shared between connectors in different groups, and Microsoft says enforcement happens in real time. A scenario that says “after the next publish” is probably wrong.
Practice: read the data policy page and block a connector in a test environment to see the error.

Create and monitor agent flows

Covers creating an agent flow, a human-in-the-loop flow, actions and connectors, input and output parameters, monitoring, and error handling. To act as a tool, a flow needs the trigger “When an agent calls the flow” and the action “Respond to the agent”, must respond in real time with the Asynchronous response toggle off, and must be published. Microsoft also states a 100-second action limit for the response.

Typical trap: picking the wrong trigger. Instant, schedule-based, and event-based triggers do not make a flow callable as an agent tool.
Practice: build one flow that returns data to an agent and one that runs a long task after responding.

Configure topics

Covers adding flows and tools to a topic, response formatting, custom prompts, custom knowledge sources, API and HTTP requests, the generative answers node, adaptive cards, and variables.

Typical trap: variable scope. A topic variable stays in its topic. To hand a value over when redirecting, you mark the receiving variable “Receive values from other topics” and map it on the redirect node. “Return values to original topics” goes the other way. A global variable cannot be converted back to a topic variable.
Practice: pass a value both ways between two topics.

Domain 2: Integrate and extend agents (40-45%)

Connect to enterprise knowledge sources

Copilot connectors, Power Platform connectors, and Azure AI Search.

Typical trap: treating all knowledge sources as the same. Each has its own authentication, limits, and refresh behavior, and those details are what scenario questions use.
Practice: add one source of each type and write down how each is authenticated and how its content stays current.

Add tools to agents

Computer use, MCP tools, tools from an existing custom connector, and REST APIs.

Typical trap: assuming MCP bypasses governance. Microsoft’s data policy page states that blocking Power Platform connectors also blocks access to tools in connected MCP servers, because they rely on those connectors.
Practice: add an MCP server, a custom connector, and a REST API tool, and compare how each is configured.

Configure multi-agent collaboration

Designing multi-agent solutions, integrating Foundry agents, existing Copilot Studio agents, and Fabric data agents, and using the A2A protocol.

Typical trap: confusing child agents and connected agents. Child agents live inside the parent agent. Connected agents are separate agents with their own orchestration. An agent that has connected agents cannot also be used as a connected agent. Microsoft also lists limits for Fabric data agents: no Redirect node, no explicit reference in instructions, and no function when the main agent is deployed to Microsoft Copilot. Microsoft marks Foundry, Fabric, and Agents SDK connections as preview.
Practice: build a main agent with one child agent and one connected agent, and note what each can and cannot do.

Integrate agents with Azure

Generative answers with Azure AI Search and Foundry, custom prompts that use the Foundry model catalog, and monitoring with Application Insights.

Typical trap: logging switches. Application Insights offers separate options, and “log everything and filter later” is not the same as not logging sensitive properties.
Practice: connect a test agent and query the telemetry yourself.

Domain 3: Test and manage agents (20-25%)

Evaluate agent performance

Creating a test set, choosing an evaluation method, and reviewing results. Microsoft’s methods include general quality, content safety, compare meaning, tool use, keyword match, text similarity, exact match, and custom. Some work only for single-response sets, and several need an expected answer.

Typical trap: choosing a method that judges the wrong thing. Tool use checks which tools or topics ran. Compare meaning judges the meaning of the answer.
Practice: run one test set with three methods and compare what each flags.

Implement application lifecycle management

Creating a solution, adding existing agents, environment variables, and Power Platform pipelines.

Typical trap: editing a managed agent in the target environment. That creates an unmanaged layer. Microsoft documents removing it through the agent’s Advanced menu and “See solution layers”.
Practice: move an agent through a pipeline across two environments.

More guides for AB-620

Next step

Use your weakest area to pick a week in the study plan. To test the areas under time pressure, try the AB-620 practice tests. Our sets are spread across these nine skill areas. The authority on scope is the AB-620 study guide.

Try the AB-620 practice exams