The short version
This sheet is a review aid, not a full course. It gathers exam facts, easily confused pairs, and decision rules for the three AB-620 domains. It does not cover everything on the exam, and Microsoft’s study guide stays the authority. Use it after you have studied, to find the gaps.
Exam facts at a glance
| Item | Fact (as Microsoft publishes it) |
|---|---|
| Exam | AB-620, Designing and Building Integrated AI Agent Solutions in Copilot Studio |
| Credential | Microsoft Certified: AI Agent Builder Associate (intermediate) |
| Time | 120 minutes, proctored, may include interactive components |
| Passing score | 700 on a scale of 1 to 1,000 |
| Questions | Number not published (Microsoft says most exams have 40 to 60) |
| Languages | 13 |
| Domain 1 | Plan and configure agent solutions, 30-35% |
| Domain 2 | Integrate and extend agents in Copilot Studio, 40-45% |
| Domain 3 | Test and manage agents, 20-25% |
| Retake | 24 hours, then 14 days; five attempts per 12 months |
| Renewal | Free yearly assessment |
Domain 1: Plan and configure (30-35%)
Key concepts
- Planning covers integration with enterprise systems, identity, channels and deployment, responsible AI, security and governance, reusable components, and internal versus external audiences.
- An agent flow used as a tool starts with the trigger When an agent calls the flow and ends with Respond to the agent.
- The response must be real time: the Asynchronous response setting is off, and the flow has to respond within the 100-second action limit. Slower work can sit after the response action.
- Add input parameters on the trigger and output parameters on the response. Publish the flow, then add it to the agent as a tool.
- Variables carry values between topics. A topic can receive values from other topics and return values to the topic that called it.
- Adaptive cards, custom prompts, the generative answers node, and HTTP requests are ways to shape or extend a response.
Confused pairs
- Agent flow versus topic: a flow runs steps and returns results; a topic controls the conversation.
- Input parameters (to the flow) versus output parameters (back to the agent).
- Topic-level variable versus a global one: a global variable cannot be converted back once it is global.
Decision rules
- If the scenario says “a person must approve a step”, think human-in-the-loop agent flow.
- If the scenario says “the flow takes too long to answer”, check whether it responds before its slow steps and stays under the limit.
- If the scenario says “a value must pass between topics”, think input and output variables.
Traps
- Leaving the asynchronous setting on for a flow that the agent calls as a tool.
- Treating planning questions as pure theory: they usually hide a security or identity detail.
Domain 2: Integrate and extend (40-45%)
This is the biggest area. Spend the most review time here.
Key concepts
- Knowledge sources: Copilot connectors, Power Platform connectors, and Azure AI Search.
- Tools: computer use, MCP tools, existing custom connectors, and REST APIs.
- Multi-agent: child agents, connected agents, Foundry agents, Fabric data agents, and agents over the A2A protocol.
- Azure: generative answers with Azure AI Search and Foundry, the Foundry model catalog for custom prompts, and Application Insights for monitoring.
Child agent versus connected agent
| Child agent | Connected agent | |
|---|---|---|
| Where it lives | Inside the main agent | A separate agent that the main agent calls |
| Own settings, authentication, deployment | No | Yes |
| Reused by other agents | No | Yes |
| Typical fit | One use case, one small team | Separate teams, separate publishing, own ALM |
More rules from Microsoft’s documentation
- Consider splitting into connected agents when the main agent has roughly more than 30 to 40 choices of action, or when tools have similar descriptions and the agent confuses them. Check your descriptions first.
- Each extra agent adds orchestration hops, so latency can rise and the testing surface grows.
- An agent that already has connected agents cannot also be a connected agent for another main agent.
- Foundry agents, Fabric data agents, and Microsoft 365 Agents SDK agents are labeled preview in the documentation. Fabric data agents cannot be a redirect target in a topic.
Decision rules
- “One task, one team, no separate publishing”: child agent.
- “Reuse across agents, own model setting, own lifecycle”: connected agent.
- “Existing REST API with no connector”: REST API tool or custom connector, depending on what the scenario already has.
- “Tools are blocked after a data policy change”: remember that blocking Power Platform connectors also blocks tools on connected MCP servers.
Domain 3: Test and manage (20-25%)
Evaluation methods
| Method | Checks | Needs expected answer? |
|---|---|---|
| General quality | Relevance, groundedness, completeness, abstention | No |
| Content safety | Harmful content by type and severity | No |
| Compare meaning | Similarity of meaning to the expected answer | Yes |
| Tool use | Whether expected tools or topics were used | Yes (tools or topics) |
| Keyword match | Any or all expected words or phrases | Yes (keywords) |
| Text similarity | Wording similarity | Yes |
| Exact match | Character-for-character match | Yes |
| Custom | Your own instructions and labels | Labels |
Confused pairs
- Compare meaning (same idea, different wording is fine) versus text similarity (wording matters) versus exact match (identical).
- Keyword match (terms present) versus text similarity (sentence construction).
- General quality can flag a correct refusal, because abstention counts. A custom method can treat a proper refusal as a pass.
Lifecycle (ALM)
- Work in a solution. Add existing agents to it. Use environment variables for values that change between environments. Use Power Platform Pipelines to move solutions.
Data policies (a common cross-domain topic)
- Connectors sit in Business, Non-business, or Blocked groups. Data cannot be shared between connectors in different groups.
- Enforcement is real time. A violation makes the Publish button unavailable.
Common traps across the exam
- Picking the most powerful option when the scenario asks for the simplest one.
- Missing a word such as “reuse”, “separately”, “approval”, or “without code” in the scenario.
- Mixing up where a setting lives: agent, topic, flow, solution, or admin center.
- Forgetting that an exam may include preview features that are commonly used.
Last-day checklist
- [ ] Re-read the three domain names and their weightings.
- [ ] Say the agent flow tool requirements out loud.
- [ ] Explain child versus connected agent in two sentences.
- [ ] List the evaluation methods and which need an expected answer.
- [ ] Review your weakest practice-exam area, not your strongest.
- [ ] Open the exam sandbox once so the interface feels familiar.
- [ ] Check your appointment, ID, and the Pearson VUE rules.
More guides for AB-620
- AB-620 Exam Format, Scoring, and Retakes
- AB-620 Study Plan: 4-Week and 8-Week
- Is AB-620 Hard? An Honest Difficulty Check
- AB-620 Prerequisites and Skills Self-Check
- AB-620 Renewal and Retakes Explained
- Is AB-620 Worth It? An Honest Look
- AB-620 vs AI-103: Which Exam First?
- Agentforce Specialist vs AB-620: Which First?
- AB-100 vs AB-620: Which Exam First?
Next step
Go through the domains explained page for the reasoning behind these points, and test yourself with the sample questions. The AB-620 overview links the rest. When you want more practice, see our AB-620 practice exams, which are original questions.