Notes from real automation work.

Failure modes, architecture decisions and practical lessons from building systems that have to keep running.

• The best AI automation habits I have came from DevOps, not AI

  Idempotency, one shared error handler, scoped credentials, alerts and a human review path. The habits that keep an AI automation alive are not AI habits.

• A workflow that only works with clean inputs is a demo

  Duplicate events, missing fields and slow APIs are normal production conditions, not edge cases. Three real failures and what they changed about the build.