Anthropic Comes for the Control Plane

Felix Rieseberg shipped a thing called Dispatch yesterday - a feature inside Claude Cowork that lets you send Claude a task from your phone, leave, and come back to finished work sitting on your desktop.

Sandboxed execution. Local files. You approve what it touches. Message from your phone. The framing is gentle, almost cozy - like giving your computer a mission before bed.

But look at what's actually been assembled here.

Cowork is a persistent agent running on your machine. Dispatch makes that agent remotely addressable. Generative surfaces handle how you interact with it. Computer use - already in the product, already capable of touching files and browsers and tools - is the execution layer. Anthropic has, in roughly the span of a product cycle, built every component of a full control plane. The announcement just didn't use those words.

The piece that isn't there yet is that it still requires your desktop to be on. Which is the kind of caveat that sounds like a limitation and is actually a countdown.

There's a version of this that's coming - you know it's coming - where the desktop requirement dissolves, where the agent isn't waiting for you to approve things in sequence, where "give Claude a mission" stops meaning a single task and starts meaning the whole day. OpenAI has been assembling the same stack from a different direction, and the race is not about who ships the best chatbot.

It's about who becomes the substrate.

One morning you'll wake up and the question won't be "which AI should I use" - it'll be whether you're on a Claude machine or not. That's the product. Everything else is just components getting bolted on in press releases.

Dispatch is a small bolt. Watch where it's going.