Anthropic just gave Claude Code a big upgrade. It’s called Claude Code Routines, and the short version is this: you can now define autonomous, multi-step AI workflows that run in the cloud on a schedule, in response to GitHub events, or via a simple API call. Your laptop can be closed. Claude keeps working.
If you run a service business and you’ve been waiting for AI agents to get practical (not just demo-worthy), this is the feature to pay attention to.
What happened
- Anthropic released Claude Code Routines in research preview. A routine is a saved configuration: a prompt, one or more GitHub repositories, and a set of connectors (Slack, Linear, Google Drive, etc.).
- Routines run on Anthropic-managed cloud infrastructure. They don’t need your machine to be on.
- Each routine supports three trigger types: Scheduled (hourly, daily, weekly, or custom cron), API (HTTP POST to a dedicated endpoint), and GitHub (reacts to PRs, pushes, issues, workflow runs).
- You can combine triggers. A single routine can run nightly and fire on every new pull request and get called from your deploy pipeline.
- Available on Pro, Max, Team, and Enterprise plans with Claude Code on the web enabled.
- You can create routines from the web UI at claude.ai/code/routines, the CLI with /schedule, or the Desktop app.
The numbers
- Minimum schedule interval: 1 hour. Anything more frequent gets rejected.
- Each routine gets its own bearer token for API triggers, scoped to that routine only.
- Runs count against your individual account’s daily run allowance. Routines are not shared with teammates.
5 things Claude Code Routines can actually do for you
- Groom your backlog automatically. A nightly routine reads new issues, applies labels, assigns owners based on the code area referenced, and posts a summary to Slack. Your team starts the day with a clean queue instead of a triage meeting.
- Triage alerts without a human staring at a terminal. Your monitoring tool hits the routine’s API endpoint when an error threshold spikes. The routine pulls the stack trace, correlates it with recent commits, and opens a draft PR with a proposed fix. On-call reviews a PR instead of starting from scratch.
- Run your own code review checklist on every PR. A GitHub trigger fires on pull_request.opened. The routine checks for security, performance, and style issues, leaves inline comments, and adds a summary so human reviewers can focus on design decisions.
- Verify deploys before anyone panics. Your CD pipeline calls the routine after each production push. It runs smoke checks, scans error logs for regressions, and posts a go or no-go to your release channel before the deploy window closes.
- Keep docs in sync with code. A weekly routine scans merged PRs, flags documentation that references changed APIs, and opens update PRs against your docs repo. No more stale docs that quietly mislead your customers for months.
The hot take
This is Anthropic quietly building the orchestration layer that most “AI agent” startups have been selling as their entire product. Routines aren’t a toy. They’re a managed execution environment with triggers, secrets, network access, and connectors to the tools teams already use. The companies charging you $500/month for a “workflow automation agent” that stitches together an LLM and a cron job should be nervous. Anthropic is bundling that into the platform. The value is shifting from “who can wire up the plumbing” to “who can write the best prompts and design the best workflows.” That’s a different skill entirely, and it favors people who understand the business problem over people who understand infrastructure.
The Agency OS play
If you run any kind of service business, here’s what to do this week. First, go to claude.ai/code/routines and create one routine. Just one. Pick the most repetitive task your team does that involves code, GitHub, or internal tooling. Backlog grooming and PR reviews are the easiest starting points because the feedback loop is fast. You’ll see results (or problems) within a day.
Second, write your prompt like a handoff document. Routines run without any human approval step. There’s no “are you sure?” dialog. That means your prompt needs to be specific about what success looks like, what the routine should never do, and where it should stop and flag a human. Vague prompts will produce vague (or destructive) results. Treat prompt-writing for routines the way you’d treat writing an SOP for a new hire on their first day.
Third, think about scope and permissions. Anything a routine does through your GitHub identity shows up as you. Commits carry your name. Slack messages come from your account. Before you hand a routine the keys to your production repo, set up a dedicated GitHub service account and connect it. Limit network access and environment variables to only what the routine needs. This isn’t paranoia. It’s basic ops hygiene for autonomous systems. Start small, watch the output for a week, then widen the scope once you trust the results.