2.6. Reflection Checkpoint

Key Takeaways

• The connection skeleton (authenticate to a project endpoint → name a deployment → send messages → read response) underlies every generative call; prefer Entra ID + managed identity over keys.
• An agent = model + instructions + tools + knowledge + thread/run; recognize when a requirement crosses from "good prompt" into "needs an agent," and place each capability at the right layer.
• Structured output shapes the response; function calling requests an action — both emit JSON, but the model never executes your code, and schema-valid is not fact-valid.
• Multi-agent = specialization + routing (coordinator or handoff); MCP connects tools, A2A connects agents, and they're complementary.
• Evaluation measures quality, not liveness — groundedness for RAG, tool-call accuracy/task adherence for agents; fine-tuning is the last lever and never adds facts.

Connecting Forward

Phase 3 turns to planning and managing the platform these solutions run on — selecting and deploying models, securing access (the Entra-vs-key choice you met here, in depth), managing cost and quota, configuring responsible-AI content safety, and wiring up monitoring and GenAIOps observability. Every deployment-type, quota, and security decision in Phase 3 directly constrains the agents and completions you just learned to build.

Self-Check Questions

• Without looking back, list the four things an agent adds to a raw chat completion, and give a one-line requirement that demands each.
• A scenario needs an agent to answer from monthly-changing policy docs in a fixed template, while looking up a customer's live balance. Name the lever or tool for each of the three needs (fresh facts, fixed format, live lookup) and justify why none of them is fine-tuning-for-facts.