Beating Cowork with Open Source Cowork

Anthropic’s release of “co-work” triggered a high-stakes scramble in the AI startup world—but one company, Camel AI, responded with a pivot that turned competitive pressure into an open-source windfall. After realizing it couldn’t effectively market a product against a major player, founder Guaal made a radical move: kill the existing product and open-source the entire system under a new name. The result was rapid traction, with a launch tweet reaching 1.7 million views, a newly purchased domain (openwork.ai), and an Apache 2 license that invited broad reuse.

Camel AI’s backstory matters because the open-source release wasn’t a last-minute wrapper around a competitor’s idea. The company had spent years building agent frameworks and multi-agent architectures designed for experimentation—especially around scaling and emergent behavior. Camel AI’s earlier work focused on testing different multi-agent designs, incorporating research patterns like task decomposition, data and instruction generation, and coordination strategies. Unlike many popular agent frameworks, the emphasis here includes pushing systems toward very large agent counts (up to a million) to study scaling laws and whether useful “emergent” behaviors appear as complexity grows.

The open-source product—described as a mainstream, desktop-oriented version of Camel AI’s multi-agent system—runs in three tiers. A React/Electron desktop front end provides the user interface across Mac and Windows. A FastAPI backend orchestrates agent execution, maintains state, and manages the workflow. The “workforce” consists of specialized agents built on Camel AI’s multi-agent core: a developer agent for code execution and technical implementation, a browser agent for information retrieval and web search, a document processing agent for writing and editing, and a multimodal agent for images and other non-text modalities.

At the heart of the architecture is a dependency-graph approach to task execution. Tasks are broken into subtasks, arranged on a directed acyclic graph (DAG), and scheduled so multiple nodes can run in parallel rather than waiting sequentially. As each subtask completes, its outputs feed dependency nodes downstream through a task-channel mechanism that assigns work to the most capable agent. The system also integrates tool use beyond a basic browser, leveraging MCPs and Playwright-based browser automation customized for navigation, retrieval, and interaction.

Model flexibility is another pillar. The system is designed to work with Frontier Labs models and has included collaborations involving Gemini 3 models, plus work with open-weight model ecosystems such as Mini Max, ZAI, and Moonshot’s Kimi K2. Still, the long-term goal is local execution: eventually, the same architecture should run on a user’s computer with local models.

Rather than selling to developers, the company frames the release as a way to “pay it forward” and attract strong agent builders. Enterprise rollout is positioned as the eventual business path, with industry-specific agent systems planned for corporate customers. In the meantime, the Apache 2-licensed code, documentation, and active GitHub contributions are meant to let others fork, customize, and build their own specialized multi-agent products.