Claude Cowork's Launch: Bringing Agent Workflows Beyond Coding

Anthropic has launched Claude Cowork for knowledge work beyond coding. I think it looks incredible, so before getting into the details I show its demonstration video.

Berman points to an explanation from Boris, the creator of Claude Code: people had been using the coding tool for non-coding tasks including vacation research, slide decks, email cleanup, subscription cancellations, recovering wedding photos, monitoring plants and controlling an oven. Those uses were surprising because the tool was designed for programmers.

With Claude Code, you give the agent access to a folder, prompt it to create code files or a directory, and iterate on the result. Berman argues that the same access can support many other kinds of work. Cowork is intended to make those workflows easier for non-developers.

Anthropic describes Cowork as an early, raw first step toward extending Claude Code to non-coding work. Berman highlights a built-in virtual machine for isolation, browser automation, data connectors and clarification questions when the agent is unsure. He welcomes isolation as a guardrail for users granting access to their computers.

At the January 2026 launch, Berman says the preview is available only on macOS and requires a Claude Max plan. He expects it to improve and become more widely available, but those are predictions about the future rather than guarantees.

In the demonstration, a user gives Cowork access to a cluttered desktop and asks it to organize the files. Cowork reads the contents and asks follow-up questions; the demonstration then shows the desktop organized into categories. Berman compares the process to Claude Code: the agent creates a plan, executes steps and verifies them.

Berman says Cowork can use existing connectors to services such as Google Drive or Box, and can control a browser to navigate websites and return when a task is complete. Users can also queue further instructions while work is in progress, a feature he values in coding agents.

The user chooses which folders to share rather than automatically granting full computer access. Berman expects the agent to ask permission before significant actions, but notes that it can still take destructive actions on a user's behalf. He urges care rather than treating the controls as a guarantee.

Berman also warns about prompt injection: an attacker can place instructions in content the agent encounters online in an attempt to change its plans. He says defenses exist but cannot provide complete protection. At the time of the video, he had seen positive early comments but had not yet tested Cowork himself.