/schedule — Cron remote agents

What it does

/schedule creates recurring cloud agents that run unattended on a cron schedule. Unlike /loop, which runs locally and requires your terminal to stay active, scheduled agents execute in Anthropic's cloud — they don't need you online, they survive restarts, and they report results back to you via notification or issue creation.

You define a cron expression (e.g., 0 2 * * * for 2 AM daily) and pass a prompt. The agent wakes up at that time, runs your task, and returns the output. It's your standing instruction to Claude in the background.

Common examples: nightly PR triage, daily codebase audits, weekly dependency checks, hourly status monitors. Any task you'd otherwise set a phone reminder for belongs on /schedule.

When to use it

Use /schedule when:

• The task runs unattended — nobody's watching, you just want the result later
• The schedule is predictable — "every night at 2 AM" or "Mondays at 9 AM", not "when this file changes"
• You want history — past runs and their outputs are logged and queryable
• The task is long-running — doesn't fit in an interactive /loop session where you're waiting

Compare to /loop:

• /loop 10m /check-tests runs locally every 10 minutes while your terminal is open — good for development
• /schedule '0 */4 * * *' /check-tests runs in the cloud every 4 hours, always, whether you're online — good for production monitoring

Real example: nightly PR triage. You schedule claude to run every night at 2 AM with a prompt like "find PRs in the backlog that are blocked or stale, assign labels, comment on them". The agent wakes up, runs the task, posts results as a GitHub issue assigned to you. You wake up to a summary instead of doing triage manually.

Try it yourself

Navigate to a project with active development. Write a prompt for a task that'd be useful to run unattended — maybe a nightly lint report, a daily dependency freshness check, or weekly documentation review. Run the /schedule skill to set it up, specifying a cron expression (use standard cron syntax: minute, hour, day of month, month, day of week). Check the scheduled agent list to confirm it was created.

Gotchas

Cron syntax is UTC, not your local timezone. 0 2 * * * is 2 AM UTC, not 2 AM your time. Adjust accordingly or hardcode UTC explicitly in your prompt.

Past runs are queryable but not re-runnable. If an agent fails, you can't replay it from the UI — you'd re-schedule or trigger it manually. Check the run history to see error messages.

Agents cannot interact with local files — they run in the cloud with no access to your working directory. If the prompt needs code context, pass file paths or snippets inline, or let the agent fetch from a public repo via git clone.

Notification defaults are quiet. Check your settings — by default you get a notification only on failure or if you explicitly asked the agent to flag something. Successful silent runs still log their output (you can query it), but you won't get paged.

Quota is shared. Scheduled agents use your Claude Code subscription quota. A nightly 10-minute task burns quota every night — budget for that. Monitor your usage if you schedule many agents.