Deploy and operate

Run proactive agents with clear triggers, scoped credentials, observable output, and human approval where it matters.

Deployment is where an agent becomes part of the team. Treat it like a small service with a narrower blast radius and a louder paper trail.

Before enabling a trigger

  • Scope credentials to the paths, repos, channels, and providers the agent actually needs.
  • Set a clear reporting channel or thread.
  • Decide which actions require approval.
  • Put a spending or runtime limit around open-ended work.
  • Test with a real event and confirm the final report is useful.

Operating loop

An unattended agent should leave enough evidence that a teammate can answer:

  1. Why did it wake up?
  2. What context did it inspect?
  3. What did it change?
  4. What does a human need to review?

If those answers are not visible in Relay, the agent is too quiet.

Promote slowly

Start in report-only mode. Let the agent recommend the action, then graduate one permission at a time: comment, label, open a PR, merge, deploy.

Design the agent contract before increasing autonomy.