Claude Cowork Just Got Serious

Claude Cowork is Anthropic’s attempt to bring the working style of coding agents to people who spend their day in documents, folders and desktop applications.

That sounds like a small product extension. It is a meaningful change in how the user has to think. Chat asks you to drive the conversation. Cowork asks for an outcome and takes responsibility for more of the path between the request and the deliverable.

What’s new

Cowork can work across local files and applications, combine information from several sources and return a finished artifact. Anthropic positions it for the high-effort, repeatable work that surrounds decisions: organising files, preparing documents, analysing material and assembling reports.

The update adds the pieces that make that practical. Computer use, so Claude can operate apps and interfaces the way a person would. Conversations that persist across devices, so a task started on the phone can be picked up on the desktop. Scheduled tasks that run without being triggered by hand. Memory that carries context between sessions instead of starting from zero.

None of those is a flashy feature. They are the boring infrastructure that separates a tool you try once from a tool you keep using.

The important feature is continuity. A person should not have to translate the same task between chat, Finder, a spreadsheet and a document one step at a time. The product tries to hold that thread while moving between the tools.

Why this update matters

I have been working this way through Claude Code and Henry at kerber.ai, where the environment is designed around agents. Cowork is interesting because it tries to make the same delegation pattern legible without asking people to become developers first.

Consider a project manager who spends two hours a day compiling status reports, or a researcher comparing a folder of PDFs. The hard part is rarely a single answer. It is collecting the material, keeping track of provenance, shaping the output and noticing what is missing. A useful agent has to participate in that workflow without making the result harder to audit.

That last condition matters. Moving through several apps increases capability and blast radius at the same time. Access should be narrow, actions should remain visible and public or irreversible steps should still require approval.

The test is delegation

Computer-use demos have existed for a while. Cowork has to prove that an ordinary person can delegate a piece of work, understand what happened and confidently take over when judgement is needed.

If Cowork succeeds, the interface for AI at work will look less like an empty chat box and more like a visible working relationship. That is a harder product problem than generating another good answer, and a more useful one.