Course 1: Run your first outcome · Chapter 3

Watch and steer the work

Read the authority that owns each fact and correct stale work before it compounds.

4 minute chapter

Chapter outline

Visibility without constant narration

Autonomy should return human attention, not consume it with progress messages. First Mate supervises through durable state and narrow wake paths, then reports outcomes, material risks and decisions.

Ask the authority that owns the question:

  • PostgreSQL: who owns the outcome and which decisions are open?
  • Kubernetes: is the intended workload healthy?
  • Herdr: what is the exact native session doing?
  • Git and its provider: what was committed, checked and opened for review?

A dashboard or event log may project all four, but the projection does not become authority.

Inspect the onboarding worker

The worker is active and First Mate has no decision to report. You attach to the named Herdr Agent and see a compatibility test failing. That terminal is direct process evidence; it does not change the Assignment or mark any message read.

Then new evidence arrives: a recent provider change already exposes the field the worker planned to add with a migration. Waiting for a polished report would waste effort, so the supervisor steers the same owned session:

Provider change 842 exposes valid_until through the current endpoint. Re-evaluate the migration before changing schema; preserve current work and report whether the existing API satisfies the outcome.

Steering changes the smallest stale assumption. It does not create a second worker to race the first or discard the work already done. If you speak directly in the attached terminal, your instruction wins over a stale brief and the supervisor reconciles the result afterward.

Look for evidence, not activity

Useful evidence answers whether the outcome is moving:

  • the same Assignment remains owned;
  • the worker incorporated the new fact;
  • the project tests now exercise the intended behavior;
  • the final artifact stays within the accepted boundary.

Token counts, terminal output and busy Pods can help diagnosis, but none proves the outcome. Next, learn how the Fleet returns the human decision without making you its polling loop.