1.5. Reflection Checkpoint

Key Takeaways

• Reason from the loop: agents perceive, reason, and act via tools, knowledge, topics, and delegation — and under generative orchestration, the model routes among components based on their descriptions, so describe everything as if the orchestrator is your reader.
• Keep the read/write line sharp: knowledge grounds answers (read-only, cited, RAG-powered); tools take actions (side effects). Requirements that mention both need both.
• Assign each protocol its job: MCP plugs an agent into resources with run-time tool discovery; A2A lets peer agents cooperate across platforms via agent cards and tasks.
• Compose, don't choose: Copilot Studio is the governed low-code front door; Foundry supplies models and pro-code agents behind it. "Integrated solutions" — the exam's own title — usually means both.

Connecting Forward

Phase 2 dives into the exam's heaviest domain (40–45%): integrating and extending agents. Every mechanism there is a concrete application of this phase — knowledge sources are RAG with different retrieval targets, MCP reappears as an attachable tool type, and multi-agent collaboration operationalizes delegation, including through A2A and Foundry.

Self-Check Questions

• A scenario requires the agent to "look up the customer's warranty status and extend it by one year." Which parts of the anatomy does each half of that sentence demand, and why can't one component satisfy both?
• Your orchestrator keeps ignoring a tool that works perfectly when invoked directly from a topic. Using the generative-orchestration model, what is the most likely cause and the first fix?
• A partner's agent runs on a non-Microsoft platform, and your agent must send it tasks. Which protocol applies, what artifact does discovery rely on, and why is MCP the wrong answer?