AB-620 plan and configure agent solutions
Plan and configure agent solutions is worth 30–35% of AB-620. It splits into three unequal parts: planning decisions you make before building, agent flows, and topic configuration. The planning part rewards judgement; the other two reward having actually built something in Copilot Studio.
Planning an agent solution
Seven planning objectives, each of which is really a design question the exam poses as a scenario:
- Integration with enterprise systems. Which mechanism reaches which system — connector, custom connector, REST API, MCP, or computer use when nothing else exists.
- Identity strategy. Who the agent acts as. The recurring distinction is whether the agent uses the user’s identity, so results respect that person’s permissions, or a single service identity that sees everything. Any scenario mentioning “users should only see what they are allowed to see” wants the former.
- Channels and deployment. Where the agent is surfaced — Teams, a website, Microsoft 365 Copilot — and what that implies for authentication and audience.
- Responsible AI strategy. Content controls, oversight, and what the agent must refuse.
- Security and governance. Data boundaries, environment strategy, who can publish.
- Reusable agent components. Build once, use in several agents.
- Internal versus external audiences. An externally facing agent needs tighter guardrails, authentication decisions and a narrower scope than an internal one.
Agent flows
An agent flow is the deterministic part of an agent — the bit that runs steps reliably rather than deciding conversationally. The objectives call out:
- Creating a flow, and configuring actions and connectors
- Input and output parameters
- Error handling
- Monitoring
- Human-in-the-loop flows, listed separately
That separate listing is a hint. Human-in-the-loop is the expected answer whenever a scenario involves an irreversible, financial, or otherwise consequential action: the agent prepares the action, a person approves, then it executes. An instruction in a prompt telling the agent to be careful is never the right answer to that requirement.
Configuring topics
Topics are where conversational behaviour is defined, and this is the longest list of objectives in the domain:
| Capability | Use it when |
|---|---|
| Agent flows in a topic | A deterministic sequence must run |
| Tools in a topic | The agent needs to call something |
| Custom prompts | You need a tailored generative response |
| Custom knowledge sources | Answers must come from specific content |
| Send HTTP request | You need to call an API directly from a topic |
| Generative answers node | The agent should answer from knowledge rather than a scripted path |
| Adaptive cards | The response needs structured, interactive UI |
| Variables | State must persist across the conversation |
| Response formatting | Output shape matters to the channel |
Adaptive cards and generative answers are the two most commonly confused. Adaptive cards are about presentation — a form, a set of buttons, a structured summary. Generative answers are about sourcing — letting the model answer from knowledge instead of following a scripted topic path.
Sample questions
Question 1. An internal HR agent must answer from policy documents, and each employee must only see policies their own permissions allow. What should the identity strategy be?
- A. A single service identity with read access to all policy documents
- B. The agent acts on behalf of the signed-in user
- C. Anonymous access with a content filter on the output
- D. A shared account for the HR department
Show answer
Answer: B
Having the agent act on behalf of the signed-in user means retrieval respects that user’s existing permissions, so no one sees documents they could not otherwise open. A single service identity with broad access would return content regardless of the individual’s rights, and neither anonymous access nor a shared account preserves per-user permissions.
Question 2. A topic must return a set of selectable options and collect two fields from the user inside the chat, rather than returning prose. Which capability fits?
- A. An adaptive card
- B. A generative answers node
- C. A variable
- D. A Send HTTP request action
Show answer
Answer: A
Adaptive cards render structured, interactive UI such as choices and input fields inside the conversation. A generative answers node produces prose from knowledge, a variable stores state rather than displaying anything, and Send HTTP request calls an API without rendering UI.
Question 3. An agent flow calls an external system that intermittently times out. Currently the whole conversation fails when this happens. What should you add?
- A. Additional input parameters on the flow
- B. A larger model for the agent
- C. An additional deployment channel
- D. Error handling in the agent flow
Show answer
Answer: D
Error handling in the agent flow lets you catch the failure and respond gracefully, which is the objective that directly addresses intermittent errors. More input parameters, a bigger model and a new channel none of them change what happens when a downstream call fails.
What to practise
Build one agent that uses a topic, an agent flow, a variable and an adaptive card together, then add a human-in-the-loop approval to the flow. That single build touches most of this domain, and the approval step is the pattern the exam returns to most often.