Claude Cowork: First General Agents from Anthropic

Entropic’s Co-work: A Detailed Overview

Key Concepts:

• Co-work: A new interface from Anthropic built on top of Claude code, designed for non-developers to interact with a local agent.
• Claude Code: Anthropic’s coding agent, providing the underlying capabilities for Co-work.
• Agentic System: An autonomous system capable of performing tasks and making decisions with minimal human intervention.
• Connectors: Integrations allowing Co-work to interact with external services like calendars and Slack.
• Skills: Pre-built or custom functionalities within Claude Code that Co-work can utilize.
• OPUS 4.5 & Sonnet 4.5: Anthropic’s language models, with OPUS 4.5 being more powerful and used for complex tasks.
• Bash Code: A command-line shell used by Claude Code to execute operations on the computer.

I. Introduction & Core Functionality

Entropic has released Co-work, positioned as “clot code for normies” – a user-friendly interface for non-developers to leverage the power of Claude Code. This represents a significant step towards a general-purpose agent capable of operating directly on a user’s computer, with access to local files. Unlike agents limited to browser interaction, Co-work can directly manipulate files and integrate with various services. It essentially provides a more accessible front-end to Claude Code’s existing capabilities, enhanced by connectors to external applications. As stated by the speaker, “this is really the first glimpse of knowledge work that we can automate with these agentic systems.”

II. Interface & Access

Currently, Co-work is available exclusively to Max subscribers within the Claude desktop application. It exists alongside the existing Chat Code interface. To utilize Co-work, users must specify a target folder for operations. Interaction occurs through a chat-like interface, similar to interacting with Claude or Chat Code. The speaker notes the rapid development of Co-work, stating it was “built in a week and a half” with “pretty much all the code was written by clot code,” highlighting the increasing utility of coding agents.

III. Desktop Organization Example & Permissions

A key demonstration involves Co-work organizing a messy desktop. The agent autonomously creates a plan, requests clarification when needed, and then executes actions with user permission. Crucially, the speaker emphasizes the need for caution, as Co-work can perform “destructive operations on your machine.” Access to folders and specific files requires explicit user granting. The example demonstrates Co-work generating a to-do list and then organizing files based on type, effectively writing and executing bash commands. The interface provides real-time progress updates and displays the specific context being utilized, offering transparency into the agent’s actions.

IV. Advanced Analysis: YouTube Transcript Project

A more complex test involved analyzing 47 YouTube video transcripts (approximately 10 hours of content) to identify themes, video lengths, and potential content directions. This task leveraged the more powerful OPUS 4.5 model. Co-work automatically detected the folder structure, identified the transcript files, and initiated dependency installation. The agent performed exploratory data analysis, generating plots illustrating topic distribution. The speaker highlights the benefit for non-developers: “you can ask it to do some really complex operations, but keep in mind since it's able to perform actions on your computer, you need to be mindful of what access you give it.”

V. Report Generation & Findings

Following the analysis, Co-work was tasked with creating a comprehensive PDF report summarizing the findings and providing recommendations. While the agent successfully created the PDF, it encountered an issue opening it within the interface. The report revealed that the primary focus of the analyzed content is on Large Language Models (LLMs), with significant coverage of infrastructure, model releases, news (around 60%), context engineering, and Retrieval-Augmented Generation (RAG) (36%). Key term analysis indicated a recent increase in content related to Gemini and agents. Recommendations included deeper exploration of agentic document workflows and multi-agent systems, aligning with the speaker’s planned content strategy. The methodology employed leveraged Claude Code’s exploratory analysis capabilities.

VI. Web Interaction & Safety Considerations

Co-work’s ability to interact with the web, even with user accounts logged in, was demonstrated by requesting the current stock prices of Tesla and Apple via Chrome. While safety nets are in place, the speaker reiterates the importance of being aware of potential security concerns due to the agent’s access to credentials and local machine operations. The agent can also utilize custom or built-in skills from Claude Code.

VII. Technical Details & Accessing Co-work

To access Co-work, Max plan users need to ensure they have the latest version of Claude Code, potentially requiring a fresh download and installation. The speaker notes current issues with connectors, which are described as “really a hit and miss,” but anticipates these will be resolved in future updates. The primary advantage of Co-work over standard chat sessions is its ability to directly access and manipulate files on the local machine. The interface includes an “Ask User Question” tool for clarification when needed.

VIII. Conclusion & Future Outlook

Co-work represents a significant advancement in accessible agentic systems, bridging the gap between powerful coding agents like Claude Code and non-developer users. While still in a research preview phase and subject to bugs, it offers a glimpse into the future of automated knowledge work. The speaker anticipates further development and competition from other companies like Google and AI, making this a crucial area to monitor. The speaker concludes by encouraging Max plan users to try Co-work and provide feedback.

Technical Terms & Explanations:

• RAG (Retrieval-Augmented Generation): A technique that combines information retrieval with language generation to improve the accuracy and relevance of LLM outputs.
• Bash: A command-line shell and scripting language commonly used in Unix-like operating systems.
• Connectors: Software components that enable different applications to communicate and exchange data.
• Skills (in Claude Code): Reusable functionalities or modules that extend the agent’s capabilities.
• UI/UX: User Interface/User Experience – the design and usability of a software application.