Teams on autopilot

A cadence agents can run, and the loop that reports it back to a human

This playbook used to be about people: form a team, align it, give it ways of working, and let it run without supervision. The same goal now applies to an agent fleet, which needs the same practices in a different order and for different reasons.

An agent fleet produces work whether or not anyone is watching, so the practices below do two things. They set the cadence the fleet runs on, and they close the loop: every run has to come back as something a human operator can check without reconstructing it from a log.

The steps form one circuit. Agree the principles, bound what the agents may do, give them a rhythm and a batch size, make them prove the work, route the result to a person, then measure what actually happened in production and feed that back into the harness. Skip the back half and you get an autonomous loop nobody can audit: plenty of output, and no way to tell whether any of it is right.

1. 01
   Principles of high performing teamsGovernance

   The principles the loop rests on, unchanged by who is doing the typing. Riskiest thing first, alignment on why, small batches, and a real answer after a failure.

2. 02
   Harness Maturity ModelGovernance

   Bound it before you automate it. An agent fleet running without governance is an unattended process holding production credentials.

3. 03
   Daily standupGovernance

   Write the cadence down instead of speaking it. Agents cannot attend a call, and yesterday’s status is the context that starts tomorrow’s run.

4. 04
   Work in small batchesCapabilities

   The unit of work. Batch size decides whether a person can meaningfully review what the fleet produced overnight, or can only wave it through.

5. 05
   Automated acceptance testing strategyCapabilities

   This is the check the agent runs on itself. Without it the only signal is the agent reporting that it worked, and that signal is worth nothing.

6. 06
   Reporting a test resultCapabilities

   A stack trace on its own leaves the operator to reconstruct the run. Give them the verdict, the environment, and what else passed.

7. 07
   Code review guidelinesGovernance

   The human gate. An agent makes a better first reviewer than a person and a worse last one, so put it first and keep the judgement call at the end.

8. 08
   Feature flagsCapabilities

   Agent-written code reaches production behind a switch, so a bad run becomes a configuration change rather than an incident.

9. 09
   Monitoring a real-world production appConnectivity

   The loop back to reality. Tests say the code does what it was asked to do; only production says whether that was the right thing to ask for.