Reference: Incident Response Agent

Architectural notes

Reference architecture for an on-call agent that triages PagerDuty alerts,
pulls context from Datadog/logs/recent deploys, drafts an initial
diagnosis, and routes to the right on-call human.

Why these picks:
- Haiku for fast triage classification; Sonnet for diagnosis writeups.
- Memory holds postmortems and "the last time this fired" context.
- GitHub MCP correlates recent deploys with alert times.
- Read-only by default; write actions (acks, rollbacks) require human approval.

Common gotchas:
- False-positive auto-responses worse than no response; bias toward
  "summarize + escalate."
- Postmortem corpus is the single biggest accuracy lever.