Anthropic's Three Tools That Work While You Sleep

Anthropic shipped three capabilities last week that together solve a problem the industry has been dancing around: how to make AI agents that actually work when you're not watching them. Scheduled tasks, Dispatch, and Computer Use aren't revolutionary on their own. Combined, they're the first infrastructure that lets you treat an AI like an employee instead of a chatbot.

The coverage has focused on features—persistent chat, remote access, GUI automation. That misses the point. What Anthropic built is an orchestration layer that happens to use your phone as the control surface and your desktop as the execution environment. You can spin up multiple parallel work sessions from a single conversation thread. Each runs independently with its own context and file access. Your phone dispatches work. Your computer executes it. You go about your day.

Nate Jones, who runs AI News & Strategy Daily, spent time testing the system and landed on a useful framing: "The distinction between work that lands on your desk and work that gets off of it is the whole game now."

That's the filter worth applying here. Most AI agent demos optimize for looking impressive. These tools optimize for making work disappear.

The Three Building Blocks

Scheduled tasks run on Anthropic's infrastructure whether your laptop is on or not. You get a repository, a schedule, and a prompt. The system executes it in a controlled cloud environment with configurable network access and environment variables. It's not revolutionary—cron jobs have existed for decades. What's different is that you configure them in natural language and they can access any MCP server you've wired into Claude.

Anthropically uses this internally to maintain code libraries. A scheduled task keeps their Go and Python codebases in sync automatically. That's production work that normally requires an engineer spending a few hours weekly on something important but never urgent. Exactly the kind of task that falls through cracks.

For non-developers, the use cases are mundane and valuable. Wake up to a digest of industry news already parsed and summarized. Get alerts when flight prices drop below your threshold. Receive reminders for bills that can't be auto-paid. Anything that happens on a schedule and doesn't require your judgment can run unattended.

The limitations are reasonable. You can't check every minute—this isn't designed for high-frequency monitoring. But hourly or daily tasks work fine, and the system integrates with existing MCP connectors without additional configuration.

Dispatch is where the orchestration layer becomes visible. The surface description—persistent chat for mobile—undersells what's happening. When you pair your phone with Claude desktop via QR code, you're not getting remote control. You're getting a dispatch layer in the literal sense.

From one conversation on your phone, you can spawn and manage multiple Claude work sessions running simultaneously on your desktop. Each session operates independently. Your phone becomes the command surface. Your desktop becomes the execution surface. The sessions run in parallel.

Pavel Hurin, a product manager, documented 48 hours using Dispatch continuously. He went to a bounce house with his kids and ran multiple work streams from the sidelines—competitor analysis, stakeholder messaging, iterative drafts. He tracked time spent: roughly 25 minutes entering commands over two days while Claude executed hours of parallel work.

"You don't have to fill in the dead time at the desk while Claude was tokenizing," Jones notes. "You can do what you need to do and go about your day and just check in with Claude when you need to."

The current constraints are real. Your desktop must be on and running. Each subtask requests folder access individually—no bulk approval yet. You can't attach files from mobile or receive output files directly. The workaround involves syncing your co-work instance to Google Drive or Dropbox. Complex multi-app tasks succeed about half the time in early testing.

Anthropically labeled this "research preview" for good reason. But the pattern it enables matters more than the polish. You're managing work, not executing it. That's the shift.

The Gap Nobody Wanted to Admit

Computer Use addresses the problem everyone knew existed but hoped would solve itself: most software will never have API connectors. MCP is excellent—it's becoming the universal standard for AI tool integration. It's also never going to achieve 50% coverage of the tools people actually use daily.

Old JIRA instances. Bespoke ERP screens. Antique SAP systems. Legacy software that runs businesses but hasn't been updated in a decade and never will be. The web is vast and most of it isn't accessible to agents through clean interfaces.

Computer Use takes the blinders off. Claude can control keyboard and mouse remotely through your desktop. You can be ten miles away and your computer will type, click, and navigate through any application you can see on screen. It's not elegant. It works.

The use case Jones highlights is one anyone who's worked in operations will recognize: "I've been the person who had to go through the agonizing process of checking two or three different antique sites for data that I then manually put into a spreadsheet and it took me like half the day."

That's the office work these tools target. Not the impressive demos. The grinding manual processes that consume hours weekly and deliver no intellectual satisfaction. The tasks you'd delegate to a junior employee if you had one.

What This Actually Means

The distinction between OpenClaw and Anthropic's approach isn't primarily about safety, though that's how it's been framed. It's about self-hosted versus managed infrastructure.

OpenClaw requires you to run your own server, configure the network, manage credentials, vet skills, and troubleshoot websocket connections. You control everything and you're responsible for everything. Anthropic's tools run on their infrastructure. You configure through natural language. The security model is managed. The tradeoff is control versus convenience.

Neither approach is wrong. They serve different needs. If you're an enterprise with specific security requirements and engineering resources, self-hosting makes sense. If you're trying to automate work without becoming a systems administrator, managed infrastructure wins.

What both approaches share is the recognition that AI agents need to work asynchronously. The chatbot paradigm—you ask, it responds, you wait—doesn't scale to real work. Real work happens in parallel. It runs overnight. It completes while you're doing something else.

Jones keeps returning to a management metaphor that clarifies the shift: "When a manager is truly managing a person, do they sit there and look over their shoulder? I mean, I've had managers that do that. I didn't like those managers."

The pattern for 2026 is treating AI agents like employees, not tools. You assign work. You check progress. You don't supervise execution. That requires infrastructure that persists across sessions, runs on schedules, and operates when you're not present.

Anthropically shipped that infrastructure last week. The tools are rough. The capabilities are real. The question is whether people will use them to get work off their desks or just create more elaborate ways to generate documents nobody needs to read.

Bob Reynolds is Senior Technology Correspondent for Buzzrag.