Claude Code Routines: Scheduled Cloud Agents, Sub-Routines and Co-Routines

Claude Code no longer needs you at the keyboard. Routines run saved agent tasks on Anthropic's cloud on a schedule, from an API call or when something happens on GitHub. Here is what they are, how to set one up, and two patterns we call sub-routines and co-routines. For the full guide, read the long article.

Three ways to start a routine

• Schedule: hourly, daily, weekdays, weekly or a one-off time, entered in your local time zone. Custom cron is available via /schedule update, with a minimum interval of one hour.
• API: a per-routine HTTPS endpoint with its own bearer token. Your alerting tool or deployment pipeline POSTs to it, optionally with context text. That text arrives labelled as untrusted, so a leaked token cannot rewrite the routine's instructions.
• GitHub: pull request or release events, filtered by branch, label, author, draft or merged state.

Set one up in two minutes

1. Open claude.ai/code/routines and choose New routine, or type /schedule in the Claude Code CLI.
2. Write a self-contained prompt: scope, allowed actions, output destination and what to do when there is nothing to do.
3. Pick repositories and an environment. Remove connectors the task does not need, because all of them are included by default.
4. Add a trigger, choose Run now to test, then read the transcript. A green badge only means the session ran, not that the task succeeded.

Sub-routines: one run, many helpers

This is our name for a pattern, not an Anthropic feature. The routine's prompt orchestrates, and it hands each step to a subagent (.claude/agents/) or a skill committed to the repository. Each helper gets its own context window and only the tools it needs, which keeps long unattended runs focused, cheaper and safer.

Co-routines: routines that hand off work

Routines cannot call each other directly, but they can cooperate through durable artefacts. For example, a nightly routine opens a draft PR, a GitHub-triggered routine reviews it, a human merges it, a merge-triggered routine updates the docs, and your deployment pipeline calls a verification routine. Use filters so no routine fires on its own output, and keep a human merge gate in the chain.

Limits worth knowing

• Daily run caps at launch were 5 on Pro, 15 on Max and 25 on Team and Enterprise, on top of normal subscription usage. One-off runs do not count.
• Routines act as you. Commits, PRs and Slack messages carry your identity. Claude pushes only to claude/ branches unless told otherwise, and never to protected ones.
• Need minute-level polling or local files? Use /loop or Desktop scheduled tasks instead.

Read next: the complete guide covers API payload security, prompt templates, subagent examples, comparisons with GitHub Actions, n8n and Trigger.dev, and a governance checklist. Sources: Anthropic announcement · Claude Code docs.