How to Use Claude Scheduled Tasks: Automate Recurring Work on a Cron

Table of contents

Open Table of contents

• What a scheduled task is
• The three places you’ll set them up
  • 1. The Claude app (for everyone)
  • 2. Claude Code routines (for people who live in the terminal)
  • 3. Managed Agents deployments (for developers building products)
• How to think about the schedule
• Patterns that actually pay off
• Setup tips from running these in production
• Claude Scheduled Tasks — 2026 FAQ
  • What are Claude scheduled tasks?
  • Do I need to be a developer to use them?
  • How is a scheduled task different from a normal Claude chat?
  • What’s a good first scheduled task?
  • Can a scheduled task take real actions, like sending emails?

What a scheduled task is

A normal Claude session is synchronous: you type, it responds, you’re there. A scheduled task is asynchronous and recurring: you define a prompt (or a whole agent workflow) plus a schedule, and Claude runs it on its own — at 7 AM every weekday, every Monday, every hour — and hands you the result when it’s done.

Under the hood it’s a cron job with an LLM at the center. You’re not writing code to glue APIs together; you’re describing the outcome in plain English and letting the agent figure out the steps each time it fires.

The three places you’ll set them up

There isn’t one button — there are three surfaces, matched to who you are.

1. The Claude app (for everyone)

The consumer Claude apps support recurring tasks: you save a prompt and a cadence, and Claude runs it on schedule and notifies you with the result. This is the no-code path — ideal for a daily briefing, a recurring research pull, a “summarize my unread newsletters every morning” job. If you’re not a developer, this is where you start.

2. Claude Code routines (for people who live in the terminal)

If you use Claude Code, you can schedule a prompt or a slash command to run on a cron cadence as a cloud agent — a “routine.” It runs server-side on your repo or workspace, so it works even when your laptop is closed. Typical uses: babysit open pull requests, run a nightly lint-and-fix pass, generate a draft post each morning for review. You define the schedule and the task; Claude Code handles the firing and the run record.

3. Managed Agents deployments (for developers building products)

For teams building on the Claude API, scheduled deployments run an agent on a recurring cron schedule — each firing spins up a session that does the work autonomously (a nightly compliance scan, a weekly report, an hourly monitor). You get a per-firing run record so you can audit successes and failures. This is the programmatic, production-grade version of the same idea.

How to think about the schedule

All three use the same mental model — what task, how often, what to do with the output:

1. The task — write it the way you’d write any good agent prompt: role, context, exact action, constraints, and a check. A scheduled task can’t ask you a clarifying question mid-run, so it must be fully specified up front. This is the single biggest difference from interactive use.
2. The cadence — daily, weekly, hourly, weekdays-only, a specific time in your timezone. Match it to how fast the underlying thing actually changes; a “daily” digest of a weekly-updated source is wasted runs.
3. The delivery — where the result lands (a notification, a file, a message, a draft). Decide this up front so the output is useful the moment it arrives.

Patterns that actually pay off

• The morning digest. “Every weekday at 7 AM, pull the latest on [topics], summarize the three things that matter, and send me a 5-bullet brief.” Replaces 20 minutes of manual scanning.
• The weekly report. “Every Monday, compile [metrics] into a one-page summary with what changed and why.” Turns a recurring chore into a review.
• The overnight worker. A coding routine that runs a long, well-specified job while you sleep — a refactor, a test sweep, a data cleanup — so you wake up to a reviewable result.
• The monitor. “Every hour, check [thing]; only message me if [condition] is true.” The best automations are mostly silent and speak up only when they matter.

Setup tips from running these in production

• Over-specify the prompt. No clarifying questions are possible mid-run. State the format, the sources, the constraints, and what to do in edge cases.
• Start with a manual test. Run the exact prompt once by hand. If it produces what you want interactively, schedule it. If it doesn’t, fix the prompt first — scheduling a bad prompt just produces bad output reliably.
• Match cadence to change-rate. Don’t run hourly against something that updates weekly.
• Keep outputs as drafts when stakes are high. For anything that goes out into the world — a published post, a sent email — have the task produce a draft for your review, not a live action. Reserve fully autonomous “just do it” for low-stakes, reversible work.
• Watch the first few runs. Scheduled jobs drift — a source changes format, a feed goes quiet. Check the early run records, then trust it.

Claude Scheduled Tasks — 2026 FAQ

What are Claude scheduled tasks?

They’re recurring jobs: you define a prompt or agent workflow plus a cron-style schedule, and Claude runs it automatically — daily, weekly, hourly — delivering the result without you being at the keyboard. They exist in the consumer Claude apps (for personal recurring prompts), in Claude Code (as cloud routines), and in the Claude API (as Managed Agents deployments).

Do I need to be a developer to use them?

No. The Claude app supports recurring tasks with no code — just a saved prompt and a cadence. Claude Code routines and Managed Agents deployments are the developer-facing versions for automating code and product workflows.

How is a scheduled task different from a normal Claude chat?

A normal chat is interactive — you’re there to answer follow-ups. A scheduled task is autonomous and recurring, so the prompt has to be fully specified up front; Claude can’t pause to ask you a question mid-run. It fires on schedule, completes the work, and hands you the result.

What’s a good first scheduled task?

A morning digest. “Every weekday at 7 AM, summarize the latest on [your topics] in five bullets.” It’s low-stakes, easy to verify, and immediately replaces a recurring manual chore — the perfect template to learn the workflow before automating anything bigger.

Can a scheduled task take real actions, like sending emails?

Yes, but be deliberate. For reversible, low-stakes work, let it act. For anything outward-facing or hard to undo, have the task produce a draft you approve rather than firing automatically — especially on unattended runs. Reversibility is the right test for how much autonomy to grant.

Related reading: The beginner’s guide to AI agents · How does Anthropic make money · How to get cited in ChatGPT answers

---

Want a system of scheduled agents running your recurring work? That’s exactly what I build — get in touch.