Claude Code's Scheduled Tasks: AI That Works While You Sleep

Anthropic just shipped scheduled tasks for Claude Code, and if you're not already neck-deep in AI automation, here's why this matters: the AI can now do things without you sitting there watching it work. Set it, forget it, wake up to results. Or at least that's the pitch.

The feature itself is straightforward—almost suspiciously so. You open the desktop app, hit the schedule tab, give your task a name and a prompt, pick your model and folder, then tell it when to run. Hourly, daily, weekly, whatever. "Boom. You now have a scheduled automation," as AI automation creator Nate Herk puts it in his walkthrough of the new feature. The scheduled task fires off, Claude Code spins up, does its thing, then stops. No babysitting required.

What makes this different from every other automation tool you've used is that it's agentic, not deterministic. That's the technical distinction, but here's what it means in practice: if Claude Code hits an error, it doesn't just die and send you a sad notification. It tries to fix itself. It attempts three different approaches, sees which one works, then updates its own process so it doesn't hit that error again. Traditional scripts follow a rigid path—step one, step two, step three. If step two breaks, everything stops. Claude Code's scheduled tasks are more like "here's what needs to happen, figure it out."

"Agentic workflows are self-healing and they can read everything in your entire project and use all your tools," Herk explains. "You are no longer the bottleneck. And these skills and these workflows can actually get better and better over time automatically."

That self-improvement loop is genuinely interesting. Herk describes a three-layer system: the agent can fix its own code when it errors, rewrite its prompt if it finds a better approach, and maintain a log file that each new session reads before starting work. So every morning at 6 a.m., his "morning coffee" task runs—it reviews his calendar, checks project status, catches him up on team updates. And theoretically, it gets better at doing that job every time it runs.

The setup is almost comically easy. Herk told Claude Code: "Take a look at my morning coffee skill. I would like to turn this into a scheduled task that goes off every morning at 6:00 a.m. Help me get this set up." Claude Code read the existing workflow, asked one clarifying question, and a minute later it was automated. That simplicity is either empowering or concerning, depending on how you feel about AI systems modifying themselves.

The Limitations Nobody Wants to Talk About

Here's where it gets messy. Your computer has to be on. The desktop app has to be open. Turn off your laptop, and your 7 a.m. automation doesn't run. Claude Code will catch up on missed tasks from the past seven days once you boot up again, but if you had something time-sensitive scheduled, tough luck.

The desktop-only requirement is temporary—Herk expects it to expand to terminal and IDE extensions soon, given how fast Anthropic ships—but right now, you're tethered to keeping that app running. For a feature marketed as "24/7 AI employee," that's a pretty significant asterisk.

Then there's the permissions problem. These tasks run without supervision, which means they need the right level of access to function but not so much access that they can wreck things. "You want to be looking at your permissions to make sure that it can't actually go off the rails and do anything like maybe make a major change to your GitHub repository or go off and delete things," Herk notes. You can configure local settings to deny certain bash commands—no deletes, no removes—but that requires knowing enough about what could go wrong to prevent it.

Each scheduled task runs in a stateless session. Every execution is fresh, with no memory of previous runs unless you explicitly build that in through log files. That design choice makes sense for some workflows—you don't want your automation accumulating context bloat—but it means you can't just assume the agent will remember what it did yesterday. If continuity matters, you have to architect it yourself.

What People Will Actually Build

The gap between what's possible and what's practical is always wider than demos suggest. Herk's morning coffee routine is genuinely useful—automated daily briefings are a clear win. But the more complex the task, the more points of failure you're introducing. And when something breaks at 3 a.m. while you're asleep, the self-healing capability better work, because you won't be there to fix it.

The notification system is bare-bones. The desktop app will alert you when tasks complete, but Herk had to set up custom hooks just to get an audible notification sound. His actual recommendation? Have Claude Code send you a ClickUp message when it finishes, because the built-in notifications aren't reliable enough to depend on.

What's interesting is the different ways people might use this. Some will treat it like a cron job that can think—scheduled report generation, data aggregation, content publishing workflows. Others will push the agentic angle harder, building systems that are supposed to improve themselves over time. The first group will probably be fine. The second group is going to discover all sorts of weird edge cases.

There's also a philosophical tension here: if you want true automation, you need to trust the system to make decisions without you. But if you're the kind of person drawn to AI automation in the first place, you probably want control over exactly how things work. Claude Code's scheduled tasks let you dial between those extremes—pure deterministic scripts on one end, fully agentic "figure it out yourself" workflows on the other—but that flexibility means you have to actually think about what level of autonomy makes sense for each task.

The question isn't whether scheduled tasks are useful—they obviously are. It's whether the agentic approach holds up once you're running real workflows with real stakes, not demos that show off the happy path. Herk is optimistic: "Now that Cloud Code is so powerful on its own, it can actually like do things in the browser as well. I truly think we're getting to that point where you can automate anything."

Maybe. Or maybe we're at the point where we can automate anything that works perfectly the first time, with enough error handling built in to catch the stuff that doesn't, assuming our computers stay on and the desktop app doesn't crash. That's still pretty useful. Just not quite 24/7.

—Zara Chen, Tech & Politics Correspondent