2.5. Reflection Checkpoint

Key Takeaways

• Choose knowledge by where truth lives and how fresh it must be: ingested unstructured content → Copilot connectors; live structured records → Power Platform connectors; owned retrieval pipelines → Azure AI Search (vector index, integrated vectorization, URL field for citations).
• Choose tools by what exists, with least-exotic-wins: existing connector → reuse it; API spec for one agent → REST import; evolving shared surface → MCP; UI only → computer use (generative orchestration required, dedicated Windows machine, credits per step).
• Scale by delegation: child agents group intents inside one solution; connected agents federate across teams and platforms — Foundry for pro-code, Fabric for data (mind its redirect and M365 Copilot limits), A2A for cross-vendor peers.
• Swap layers, not platforms: node-level Azure OpenAI grounding outranks agent knowledge; prompts pick catalog models per task; Application Insights flows only after you connect it.

Connecting Forward

Phase 3 moves from extending agents to planning and configuring them — the 30–35% domain. The integration options you now know become planning constraints: identity strategy decides what knowledge can honor per-user permissions, channel strategy decides who can reach the agent, and topics/flows give deterministic structure to everything the orchestrator routes.

Self-Check Questions

• A scenario needs live inventory numbers, searchable third-party wiki content, and a VNet-isolated legal corpus — name the three knowledge sources and the single property of each requirement that decides it.
• Rank the four tool types for "a stable HR API already published as a governed custom connector, used by three Power Apps" — and defend why the newest technology is the wrong answer.
• Your parent agent must delegate to a partner company's logistics agent and to your own Fabric data agent. For each: which mechanism, and which constraint (trust boundary vs. invocation limits) shapes the design?
• Telemetry must prove that answer latency degraded after last week's knowledge change. Which monitoring surface, what prerequisite, and what query language?