Claude Code routines are interesting because they make agent work feel less like a chat session and more like automation. That is the shape I expect serious coding agents to take: scheduled, event-driven, scoped, and reviewable.
Anthropic’s Claude Code week 20 notes describe routines as a way to run repeatable agent work from triggers and schedules. That is a different mental model from “ask the assistant to help me with this file.” It is closer to “run this workflow when the conditions make sense, then show me what changed.”
That matters because a lot of engineering work is repetitive but not quite mechanical. Dependency cleanup, flaky-test investigation, documentation drift, issue triage, small migrations, and release checks often need judgment but also benefit from a fixed route. A routine can carry the prompt, constraints, and expected output instead of relying on somebody to remember the perfect instruction every time.
The risk is the same as with CI: bad automation creates very consistent mess. Routines need tight scopes, clear outputs, and review gates. Used well, though, they push agents into a healthier place. Less magic chat window, more boring engineering system. I mean that as a compliment.