Pro Plan from January 16, Team and Enterprise Plan Expansion from January 23
Anthropic unveiled Cowork, a new function that extends Claude beyond a simple conversational tool to a subject that actually performs work.
The core of Cowork is that it provides an agent-type work environment that directly accesses the user's computer's local folders to read, edit, and create files, and plans and executes tasks. It is evaluated as an attempt symbolizing that generative AI has entered the stage of actually taking charge of and handling work, going beyond 'advisor' or 'assistant.'
Cowork's biggest characteristic is autonomy. While existing chat-based AI responded to users' questions, with Cowork, when task instructions are received, Claude independently establishes plans, executes step by step, and shares progress and results with users. This is a structure that extends the plan→execute→feedback loop verified in Claude Code to general work such as document tasks and file management. As a result, users can handle complex tasks with just goal-level instructions rather than repeatedly inputting detailed commands.
Also, Cowork can be linked with existing connectors provided, and when combined with the Chrome environment, tasks requiring browser access can also be performed. This suggests the possibility of expanding into a hybrid work agent that simultaneously handles local file systems and web tasks.
However, Anthropic also relatively clearly mentioned Cowork's risks and limitations. Since Cowork can actually edit and delete files, it can perform potentially destructive tasks. While Claude basically requires user confirmation before important tasks, depending on commands, file deletion and mass editing are also possible. The risk of prompt injection that can occur in the process of handling external web content was also specified. Anthropic explained that it has built defenses against this but drew a line that agent safety is still a developing area across the industry.
Cowork's unveiling is interpreted as a strategic turning point where Anthropic is trying to position Claude not as an 'AI tool' but as an 'AI colleague.' While generative AI competition has been centered on model performance and benchmarks, Cowork is the stage of testing how deeply AI can become involved in actual work environments. This shows that the competitive axis is shifting from competition to make smarter models to who provides more actual work productivity.
Anthropic defined Cowork as a research preview and announced the policy of rapidly improving functions through user experiments. Windows support, sync between devices, and safety strengthening in the future were also announced. Claude Max users can immediately use Cowork in the macOS app, and users on other plans can register on the waitlist to receive access permissions sequentially.
This Cowork launch clearly reveals the flow of generative AI moving from 'tools that speak well' to 'entities that work on behalf.' Now that AI has begun to enter deeply into users' file systems and work contexts beyond code and documents, the next phase of generative AI competition is moving to 'how safely and how much actual work can be entrusted' rather than model performance.
