Claude's Scheduled Tasks: AI Automation That Runs

The fantasy of autonomous AI agents has produced more vaporware than working software. We've watched projects promise self-running systems that mostly delivered sophisticated ways to crash in the background. So when Anthropic quietly added scheduled tasks to Claude Co-work last week, I approached it the way I approach all automation promises: with curiosity and skepticism in equal measure.

Kenny Liao, who builds tools for AI workflows, spent the past few days stress-testing this feature. His results suggest something genuinely useful—with caveats that matter.

What Actually Works

The concept is straightforward: you write a prompt, set a schedule, and Claude runs it without you. Liao demonstrates three workflows he now runs daily. His morning starts with an AI news brief already waiting, pulled from sources he's configured. Before he opens his email client, Claude has already triaged his inbox by priority. Every Monday, a Stripe sales analysis sits ready in his documents folder.

These aren't flashy examples. They're boring in the way reliable plumbing is boring—you notice when it's not there.

The most interesting use case involves what Liao calls a "brain dump" file. Throughout the day, he voice-records ideas, facts, and reminders into this single document. Every night at a scheduled time, Claude reads this growing file and distributes its contents to appropriate destination files: project notes, personal context, video ideas, content for his Substack. The system maintains itself without manual filing.

"I don't have to plug it in using hooks every time I talk to Claude," Liao explains. "Now it just runs on an automated schedule without me ever noticing it in the background."

This addresses a real problem with AI assistants: context decay. Every conversation starts fresh unless you manually feed the system what it knew yesterday. Automated context maintenance could be more valuable than any individual scheduled task.

The Constraints

Here's what Anthropic's documentation downplays: scheduled tasks in Claude Co-work only run when your computer is on and the Claude app is open. Turn off your machine or quit the app, and your scheduled tasks sit idle. Co-work attempts to catch up on missed tasks when you reopen, but this isn't guaranteed.

For a feature marketed around autonomy, that's a significant limitation. The promise is AI agents working while you sleep. The reality is AI agents working while your computer is awake and running specific software.

Liao built a workaround—a plugin for Claude Code that uses your operating system's native cron scheduling. This runs whether or not the Claude app is open, which is how scheduled tasks should work in the first place. That he needed to build this suggests the official implementation isn't finished.

There's also the question of reliability. Liao demonstrates successful runs, but scheduled automation fails in predictable ways: API timeouts, authentication issues, malformed prompts that work manually but break in automation, rate limits hit during overnight processing. None of these appear in his video, which either means he edited them out or hasn't run the system long enough to hit them.

What This Enables

The practical applications extend beyond Liao's examples. Automated competitor research that runs weekly and compiles changes. Nightly file organization that keeps project directories clean. Team updates distributed through Slack without anyone remembering to send them. Error monitoring that checks production systems and surfaces issues before your team arrives.

These workflows share a pattern: they're tasks humans do inconsistently because they're boring and easy to forget. The value isn't that AI does them better—it's that AI does them at all.

Liao notes you can specify which Claude model runs each task. Routine summaries can use the faster Sonnet 4.6, saving the expensive Opus runs for complex analysis. This granular control suggests Anthropic is thinking about costs at scale, which matters if scheduled tasks become standard practice.

The Larger Pattern

Scheduled tasks represent a shift in how we're meant to use AI tools. Instead of conversational interfaces where you prompt and wait, we're moving toward systems you configure once and monitor periodically. The AI becomes infrastructure rather than assistant.

This creates new problems. How do you debug a prompt that works in conversation but fails in automation? How do you know when a scheduled task has drifted from useful to useless? Who reviews the nightly file reorganization to ensure the AI hasn't developed a bizarre categorization scheme?

Liao's demonstration sidesteps these questions, but they matter. Automation that runs unsupervised needs different safeguards than interactive AI. We don't yet have established practices for this.

Anthroptic's suggested use cases—daily briefings, weekly reports, recurring research, file organization, team updates—are deliberately conservative. Nothing here scrapes the edge of what's possible. That's probably wise. The gap between a working demo and reliable production automation is where most AI projects die.

What's Missing

The technical implementation raises questions Liao's demonstration doesn't address. Scheduled tasks in Co-work use Claude Code under the hood, which means they can access MCP servers and custom skills. But there's no discussion of error handling, retry logic, or notification when tasks fail. The plugin he built for Claude Code sends desktop notifications, which is something, but falls short of a proper monitoring system.

Cost transparency is absent. If you schedule hourly tasks across multiple workflows, all running Sonnet or Opus, the charges accumulate quietly. There's no dashboard showing scheduled task usage, no way to set spending limits specifically for automation. This will surprise someone.

Security and permissions deserve more attention than they receive. Scheduled tasks can access your email, calendar, file system, and external APIs—all without you present. Liao mentions you can specify "only the needed permissions," but the video doesn't demonstrate how granular this control actually is.

Whether This Matters

Automation features tend to divide into two categories: those that become invisible infrastructure and those that remain perpetual science projects requiring constant maintenance. Scheduled tasks could go either way.

The determining factor will be reliability. If these tasks run predictably for weeks without intervention, people will build dependencies on them. If they require frequent debugging and adjustment, they'll join the long list of automation attempts that worked great in demos.

Liao's enthusiasm is genuine—he's using this daily. But he's also someone who builds AI tools for a living, with the debugging skills and patience that implies. Whether scheduled tasks work for people who just want their morning email sorted remains to be seen.

The feature is live now for Claude Co-work users on Mac, with Windows availability unclear. Liao's plugin for Claude Code is open source on GitHub. Both approaches let you test whether autonomous AI agents are ready to handle your boring tasks.

Just keep your computer on.

Bob Reynolds has covered technology since 1974 and remembers when automation meant a cron job you wrote yourself.