3.4. Reflection Checkpoint

Key Takeaways

• Plan from audience and consequences: identity decides per-user data exposure (Entra ID user context vs. maker credentials vs. none), channels go where the audience already is and interlock with auth, and responsible AI is four authored layers — content, transparency, oversight, governance (DLP, environments, admin center).
• Build for reuse: solution-packaged topics, flows, prompts, and templates make the second agent an assembly job; conflicting internal/external audience columns argue for two agents sharing components.
• Flows are authored determinism: "when an agent calls the flow" makes a tool; parameters are the reasoning↔procedure contract; approvals checkpoint irreversibility; run-after conditions are error handling you must author; capacity is a reliability budget.
• Topics have two faces — selection (description/trigger phrases) and execution (nodes) — and variables are scoped memory: topic by default, global by promotion, moved explicitly via redirect wiring and parameters.

Connecting Forward

Phase 4 closes the build loop: proving the agent works (test sets, evaluation methods) and shipping it like software (solutions, environment variables, pipelines). Everything you configured in Phases 2–3 becomes cargo — packaged, parameterized, and promoted dev → test → prod.

Self-Check Questions

• A case study serves logged-in employees in Teams and anonymous customers on the web, with Dataverse-backed personal answers for the first group only. Derive the identity + channel + knowledge design, and name the requirement that forbids one-agent-one-auth.
• The orchestrator invokes your refund flow but never collects the order ID first, and the topic can't read the outcome. Which two halves of the parameter contract are broken?
• Answers in the compliance topic must come only from the compliance library, formally worded, with citations — while the rest of the agent stays casual. Which node, which three settings, and which priority rule make this work?