Scheduled agents

Every way you’ve run Claude so far has one thing in common: you started it. A prompt, a command, a script you ran. Scheduled agents break that last tie.

A scheduled agent (a “routine”) is a task you hand to Claude along with a time — and Claude runs it on that schedule, on its own, without you there.

“Every weekday at 8am, summarize what changed in the repo overnight and what needs my attention.” You set it once. From then on, the summary is waiting for you with your coffee — and you never touched the keyboard to make it.

Setting one up

Inside a session, the slash command kicks it off:

/schedule

You describe the job in plain English and say when — “every Monday at 9am,” “the first of each month,” “tomorrow at 3pm just once.” Claude turns that into a routine and registers it. From then on it fires on schedule, runs the task headlessly (the engine from Lesson 11.6, now on a timer), and leaves you the result.

You can list what you’ve scheduled, change a routine’s timing, or delete it — the same way you’d manage any standing reminder.

What it’s actually good for

The sweet spot is recurring chores you’d forget or resent doing by hand:

• A morning digest of overnight commits, new issues, failed builds — triaged before you’re online.
• A weekly dependency check — “any of our packages have security updates? open a summary.”
• A monthly cleanup pass — stale branches, dead TODOs, docs that drifted from the code.
• A one-time nudge — “tomorrow at 2pm, remind me to follow up on the migration, and check whether it finished cleanly.”

Notice these aren’t urgent. They’re important-but-easy-to-drop. A schedule is how you stop dropping them.

The thing to keep in mind

A scheduled agent runs when nobody’s watching — which is the whole point, and also the whole risk. The Level 6 instinct still holds, just shifted earlier in time:

You’re approving the blast radius now, for every future run, instead of in the moment. So scope the job narrowly and lean toward ones that report rather than ones that change things unsupervised.

A routine that summarizes and pings you? Delightful, low-stakes. A routine that force-pushes to production at 3am with nobody awake? That’s how you wake up to a bad morning. Start with read-and-report routines; graduate to action only once you trust the shape of the job.

What’s next

A clock is one kind of trigger. The other big one is an event — a pull request opening, an issue being filed. The final lesson puts Claude into your team’s pipeline, where it reacts to those the moment they happen: Claude in CI.