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:
- Why did it wake up?
- What context did it inspect?
- What did it change?
- 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.