Claude Code’s recent direction is a useful signal for where coding agents are going. The category started with “ask an assistant to edit code”. It is moving toward “define a routine and let it run when something happens”. That is a much bigger shift, because it turns the agent from a helper into part of the automation layer.
Anthropic’s Claude Code week 16 notes describe routines on the web: define the prompt, repos, and connectors once, then trigger the routine from events such as a PR opening, a release being published, a webhook, or a schedule. That is not just a nicer CLI. It is agentic automation tied to normal software events.
The useful version of this is obvious. Run a dependency review when a package changes. Prepare release notes when a tag is created. Check documentation after an API change. Triage a queue of small issues. The dangerous version is also obvious: a vague prompt with broad repo access quietly doing work nobody reviews properly.
So the engineering problem becomes less about whether the agent is clever and more about whether the routine is bounded. What can it touch? What evidence does it need? When should it open a PR instead of pushing? Who reviews the output? These are boring questions, but they are the questions that decide whether agent automation helps or becomes another production risk.