AB-620 integrate and extend agents explained
Integrate and extend agents in Copilot Studio is worth 40–45% of AB-620 — the largest domain and the one that distinguishes this exam from every other agent certification. It is a catalogue of mechanisms for connecting an agent to the outside world, and the exam’s favourite question shape is: here is a requirement, which mechanism fits?
Learn it as a decision table, not as a list.
The decision table
| Requirement | Mechanism |
|---|---|
| Reach a common SaaS system with a ready-made integration | A Copilot connector or Power Platform connector |
| Reach an internal service that has a documented API | A REST API tool, or a custom connector |
| Expose tools to the agent through an open, model-facing standard | MCP tools |
| Search your own indexed content for grounding | Azure AI Search |
| Reach a system with no API at all | Computer use |
| Query enterprise data in Fabric | A Fabric data agent |
| Hand work to another agent on another platform | The A2A protocol |
| Hand work to a specialist agent you built in Foundry | Foundry agent integration |
If you can reproduce that table from memory, you have most of this domain.
Enterprise knowledge sources
Three objectives: Copilot connectors, Power Platform connectors, and Azure AI Search. Knowledge sources are about what the agent can know; tools are about what it can do. Questions that describe answering from documents or records want a knowledge source; questions that describe performing an action want a tool.
Enterprise systems named in Microsoft’s audience profile — ServiceNow, SAP and others — are reached through connectors rather than bespoke code.
Tools
- Custom connector as a tool. You already have a connector; surface it to the agent.
- REST API. Add an API directly.
- MCP tools. Model Context Protocol is a standard way to expose tools to models. Reach for it when tools should be defined once and consumed by models generally, rather than wired per-agent.
- Computer use. The agent drives an application interface directly. It is the answer only when there is no API and no connector — it is slower and more brittle, so the exam expects it as a last resort. The objective also says monitor computer use, which is a hint that observing what it does matters.
Multi-agent collaboration
Five objectives here, and the distinctions matter:
- Design multi-agent solutions — decomposing a problem across specialised agents.
- Integrate a Foundry agent — a code-first agent built in Foundry, called from Copilot Studio.
- Integrate an existing Copilot Studio agent — reuse inside the same platform.
- Integrate a Fabric data agent — for enterprise data questions.
- A2A protocol — the open standard for agent-to-agent communication, which is what you use when the other agent is not yours and not on your platform.
The sorting rule: inside Copilot Studio → integrate an existing agent. Your own code-first agent → Foundry integration. Someone else’s agent, elsewhere → A2A.
Integrating with Azure
- Generative answers using Azure AI Search with Foundry — grounding at enterprise scale.
- Custom prompts against the Foundry model catalog — when the default model is not the right one.
- Monitoring with Application Insights — the answer to any question about telemetry, traces or diagnosing agent behaviour in production.
Sample questions
Question 1. A support agent must answer questions from incident records in ServiceNow, which your organisation already licenses. What is the most appropriate approach?
- A. Connect the agent to ServiceNow through a connector as a knowledge source
- B. Use computer use to operate the ServiceNow web interface
- C. Copy incident records into Dataverse nightly and query that
- D. Export incidents to a document and upload it as a file
Show answer
Answer: A
A connector provides a supported, maintained integration with a common enterprise system, which is exactly the scenario Microsoft’s outline describes for systems such as ServiceNow and SAP. Computer use is for systems with no API, rebuilding the data in Dataverse duplicates a source of truth, and a static document export goes stale immediately.
Question 2. Your Copilot Studio agent must delegate specialist analysis to an agent maintained by a partner organisation, running on a different vendor's platform. Which approach fits?
- A. Integrate it as an existing Copilot Studio agent
- B. Add it as a Fabric data agent
- C. Use the A2A protocol
- D. Render its output in an adaptive card
Show answer
Answer: C
The A2A protocol is the open agent-to-agent standard, which is what is needed when the other agent is on a different platform and not yours to modify. Integrating an existing Copilot Studio agent assumes it is in your platform, a Fabric data agent queries data rather than delegating analysis, and an adaptive card renders UI.
Question 3. Agents in production sometimes produce unexpected results, and your team needs traces and telemetry to diagnose what happened. What should you configure?
- A. Environment variables for each agent
- B. Monitoring through Application Insights
- C. A larger test set
- D. Adaptive cards on every response
Show answer
Answer: B
Application Insights is the monitoring objective in this domain and provides the traces and telemetry needed to diagnose behaviour after the fact. Environment variables hold configuration, a larger test set helps before release rather than in production, and adaptive cards affect presentation.
What to practise
Configure each row of the decision table at least once, even trivially. The exam does not ask you to recall a menu path — it asks which mechanism a requirement calls for, and that judgement comes much faster once you have seen what each one takes to set up.