You've been using Claude Code — Anthropic's AI coding assistant — the way you use a calculator. Open it, type a prompt, wait, close the tab. The tool sits quiet until you poke it. That's the whole interaction model for every chat-based AI since 2022: you ask, it answers, the lights go off. Good for questions. Fine for drafting a function. Useless for the half of engineering work nobody wants to do — nightly triage, stale-branch cleanup, dependency bumps, the 11 p.m. "why did CI go red" check.

That half doesn't fit the shape of a chat agent. You can't prompt what you forgot to remember.

On April 14, 2026, Anthropic quietly shipped the fix. It's called Routines, and it's the first time an AI coding agent wakes itself up.

What actually shipped

A Routine is a saved Claude Code setup — a prompt, one or more GitHub repos, and a bundle of connectors (Slack, Linear, whatever) — that fires on its own. Anthropic's words, not mine: "packaged once and run automatically." Three triggers, and one Routine can use all three at once:

  • Scheduled — cron (the Unix recipe for "run this at 3 a.m. every Tuesday"), minimum interval of one hour.
  • API — a webhook, meaning any other system can POST to a URL and wake the agent.
  • GitHub — pull-request and release events, filtered by branch, label, author, draft state, whatever.

The twist: Routines don't run on your laptop. They run on Anthropic-managed cloud infrastructure, as full Claude Code sessions, "so they keep working when your laptop is closed" (Anthropic docs). No approval prompts mid-run. No permission-mode picker. Set it once, it runs forever 😼

The catch Anthropic wrote down itself

Read this line carefully, because it's the whole story: "Anything a routine does through your connected GitHub identity or connectors appears as you: commits and pull requests carry your GitHub user, and Slack messages, Linear tickets, or other connector actions use your linked accounts."

Translation: the agent impersonates you. When a 3:47 a.m. Routine opens a bad PR, the blame column on GitHub says your name. When it pings #eng-oncall on Slack, it's your avatar. There is no "agent" identity in any enterprise IAM system — Okta, Entra, AWS SSO — that currently understands "this action was performed by a cron job made of English sentences."

The Register's Thomas Claburn called Routines "mildly clever cron jobs" sitting between scheduled tasks and true autonomous agents. Accurate, and also the problem. They're clever enough to ship code, dumb enough to need supervision, and structured so the supervision line runs through your personal GitHub account.

The economics nobody's ready for

Anthropic set daily caps for the preview, per SiliconANGLE and 9to5Mac: 5 runs/day on Pro, 15 on Max, 25 on Team and Enterprise. Above that, metered overage billing kicks in. Routines eat subscription usage the same way interactive sessions do — except you're not the one deciding when they fire. A noisy GitHub repo can burn your entire daily quota by 9 a.m., and Anthropic silently drops webhook events past the hourly cap. You won't know a Routine missed its trigger unless you're reading logs 😹

Branch safety defaults are sane: Claude can only push to branches prefixed claude/ unless you flip "Allow unrestricted branch pushes." Most teams will flip it within a week. That's how these things go.

What this means if you ship code

Routines is the first time "AI that works while you sleep" is a line item on a real invoice, not a demo. The question stopped being can an agent do this? and became who owns the pager when the agent is the one paging?

Finance doesn't have a category for it. Security doesn't have a review template. On-call rotations don't have a slot for a teammate that's a cron expression. Every playbook your company wrote before April 14 assumed a human pressed the button.

Claude Code used to be a tool you invoked. After April 14, it's a worker you hired 🐈‍⬛