Content plan
Learners can trust output they didn't write, and debug production with AI.
Outcome
Learners can trust output they didn’t write, and debug production with AI.
Planned coverage
- Verification in practice: review for intent and coverage, not line-by-line authorship.
- The sensor net as the gate.
- Spotting plausible-but-wrong output.
- The “agent grades its own homework” guardrail made concrete: humans own the failure model and the critical acceptance criteria.
- Debugging in production with AI: point an agent at the observability built in Part 2.
- MCP examples: Postgres to query state, and MCP against logs or traces to investigate incidents.
- Deep code-review discipline remains its own later course.
Agent payoff
The AI is a debugging copilot bounded by the telemetry you instrumented.
The full loop, closed
Bounded spec → agent implements in the harness → sensor gate → ship → observe → AI-assisted debug → back around.