Anthropic、業務全般向け新ツール「Cowork」を発表 Introducing Cowork: Claude Code for the rest of your work

Anthropic has unveiled Cowork, a new product that aims to bring the agentic experience of Claude Code to work beyond software engineering. The pitch, as the name suggests, is to let knowledge workers delegate everyday tasks to a Claude-powered agent in much the same way developers already hand off coding chores to Claude Code.

Claude Code has become one of Anthropic's most talked-about offerings of the past year. Running in the terminal, it can read and edit files, execute commands, and autonomously drive multi-step coding tasks for extended periods. Developers have embraced it as a way to offload boilerplate work, refactors, and exploratory debugging. Cowork appears to take that same delegation pattern, where the user sets a goal and the agent grinds through the steps, and apply it to general office work such as research, document drafting, data wrangling, and routine operational tasks.

The move fits a broader industry shift from chat-style copilots toward agents that can actually execute multi-step workflows on a user's behalf. OpenAI has pushed in this direction with ChatGPT Agent and its Operator-style browsing capabilities; Google is weaving Gemini deeper into Workspace; Microsoft offers Copilot Studio for building enterprise agents; and startups like Cognition's Devin and Manus are racing to define what an autonomous knowledge worker should look like. Anthropic itself already exposes Computer Use through its API and sells Claude for Work to enterprises, so Cowork looks like a natural extension that packages the tool-use, planning, and long-horizon context handling honed in Claude Code into a form non-developers can adopt.

Details released so far are limited, and key questions remain open: how Cowork is priced, which surfaces it runs on, how it integrates with Claude for Enterprise and existing SaaS connectors, and what guardrails Anthropic puts in place for sensitive business data. Given Claude's reputation for careful instruction following and its constitutional-AI safety framing, enterprise buyers will likely scrutinize how Cowork handles permissions, audit trails, and human-in-the-loop checkpoints.

If Cowork delivers even a fraction of the productivity gains Claude Code has shown for engineers, it could meaningfully reshape how routine white-collar work is structured. That outcome is far from guaranteed, since general office tasks tend to be messier and less verifiable than code, but Anthropic is clearly betting that the agentic playbook generalizes. More concrete details will be needed before the market can judge how directly Cowork challenges incumbents like Microsoft Copilot or the emerging crop of vertical AI agents.