Claude Cowork: Anthropic's New AI Agent That Runs Code on Your Files (Max Feature)
Quick Overview
Anthropic launched Claude Cowork, a new feature for Max subscribers that allows parallel AI agents to work on local files, essentially wrapping Claude code for non-developers to perform tasks like organizing screenshots by year and month, culminating in an interactive artifact summarizing the work done.
Key Points: Claude Cowork is a new research preview feature available to Anthropic's MAX subscribers, designed to let parallel agents work on local files and tools. The feature operates by creating a secure sandbox on the user's computer where Claude code can run to perform file operations, such as organizing 5,104 screenshots from 2023-2025 into monthly folders. The agent successfully executed complex steps, including running shell commands, despite an initial command failure due to output size limits, which it then self-corrected. A key capability is creating an interactive artifact (a JSX file) at the end of the task, which summarizes statistics like total screenshots per year/month and allows user interaction, as demonstrated with yearly/monthly distribution charts. The process involves defining a task, providing context (the folder location), and allowing the agent to generate and run necessary scripts, all accessible via a simple UI instead of the command line. The presenter notes that this feature bridges the gap for non-developers who want to leverage the power of Claude code execution without needing CLI expertise.
Context: The video introduces and demonstrates Claude Cowork, a new capability for users subscribed to Anthropic's MAX tier. This feature allows the AI agent to execute tasks that interact directly with the user's local file system, a significant expansion beyond typical chat-based interactions. The demonstration focuses on a file organization task: sorting a large collection of screenshots from multiple years into organized monthly folders.
Detailed Analysis
Anthropic has released Claude Cowork, an early research preview feature exclusively for MAX subscribers, enabling AI agents to run code and operate on local files. The presenter subscribed to MAX specifically to review this feature, which functions by creating a secure sandbox on the user's machine. The demonstration involves organizing screenshots; the user initially prompts the agent to sort screenshots from 2024 and 2025 into monthly folders. The agent plans out the steps, which include file exploration, script creation (based on modification date), and execution, even showing the underlying shell commands being run. An initial command to list files failed because the output was too large (799.4KB), but the agent recognized the failure and proceeded to adjust its approach, eventually moving the 2024 and 2025 files. When asked about 2023 files, the agent intelligently integrated that request into the existing plan. Upon completion of the file organization steps, the agent was prompted to create an artifact summarizing the work. This artifact, a JSX file, rendered an interactive dashboard showing statistics: 5,104 total screenshots across 2023 (34), 2024 (2,583), and 2025 (2,497). The interactive artifact allowed toggling between yearly data via bar or line charts, showing monthly distributions, effectively delivering a comprehensive, visual report of the task's results, which the presenter found impressive.