Operations is where agents earn their keep and where they page you at 3 a.m.
Use them to draft and diff, not to own production, until the promotion checklist is real.
Pack
- Runbook (what “green” looks like)
- Last incident note
- Architecture sketch you actually maintain
- Access: staging credentials only
Spec
- Deliverable: a change in a branch, a checklist executed, or a report
- Forbidden: production applies, DNS, IAM, customer mail
- Critic: tests, health checks, or a second pair of eyes
Loop
- Plan against the runbook
- Act in staging / a copy
- Critique with tests
- Human promote
If there is no staging, there is no agent. There is a person with a checklist.
Incidents
Freeze autonomous apply during an incident. Use chat to summarize logs you already pulled, not to “fix prod.” Afterward, improve the runbook — that is the pack getting smarter, which is the whole point of memory as files.
Related
- Loops
- Eval
- Agents in 2026 (repo agents + tests)