# Eigent AI taking on Claude Cowork

Source: https://www.youtube.com/watch?v=-UoxWCsqIa0
Recap page: https://rapidrecap.app/video/-UoxWCsqIa0
Generated: 2026-01-21T14:07:58.52+00:00

---
## Quick Overview

Startup Camel AI, originally developing the multi-agent system product Aent AI, made the radical decision to open source their entire codebase under the name OpenWork.ai, complete with an Apache 2 license, immediately following Anthropic's announcement of Claude Cowork to avoid direct competition.

**Key Points:**
- Camel AI's founder, Guhao, decided to kill the Aent AI product and open source it after Anthropic released Cowork, realizing they could not market against a large company like Anthropic.
- The open-sourced project was immediately renamed to openwork.ai and released under an Apache 2 license, gaining significant traction with one tweet reaching 1.7 million views.
- The core technology, built on Camel AI's multi-agent framework research, focuses on building hyper-specialized, customizable multi-agent architectures capable of scaling up to a million agents.
- The OpenWork architecture utilizes three tiers: an Electron desktop front end, a FastAPI agent backend for orchestration, and the agents themselves built on the Camel AI multi-agent core.
- The system employs task decomposition using a Directed Acyclic Graph (DAG) structure, allowing subtasks to run in parallel, coordinated by a coordinator agent.
- It includes four pre-built specialized agents: a developer agent with file system access, a browser agent, a document processing agent, and a multimodal agent.
- Camel AI's long-term goal for the open-sourced system is for it to eventually run entirely with local models, though current iterations work with proprietary models like Gemini 3.

**Context:** The video details the strategic pivot taken by the startup Camel AI, founded by Guhao, in response to Anthropic launching its 'Cowork' feature, which directly competed with Camel AI's product, Aent AI. Camel AI has a history dating back almost three years, initially focusing on academic research into training models for action-taking, later pivoting to build agent frameworks and systems designed to test multi-agent architectures, leading up to the now open-sourced Aent AI software.

## Detailed Analysis

Camel AI's Aent AI, a software product designed to run multi-agent systems, was open-sourced as OpenWork.ai with an Apache 2 license immediately after Anthropic released Cowork, a move that generated substantial community interest. The underlying technology stems from Camel AI's foundational research into agent frameworks, aiming to test emergent behaviors in massive multi-agent systems (up to a million agents) by incorporating patterns like Chain of Thought and coordination mechanisms. The OpenWork architecture is structured in three tiers: the React Electron desktop application (front end), a FastAPI backend for state maintenance and orchestration, and the agent workforce utilizing the multi-agent core. Key to its functionality is task decomposition managed via a Directed Acyclic Graph (DAG) that enables parallel processing of subtasks assigned to specialized agents. These specialized agents include a developer agent (capable of file system interaction, demonstrated by organizing a desktop), a browser agent, a document processing agent, and a multimodal agent. Although currently compatible with proprietary models like Gemini 3 and models from companies like MiniMax and ZAI, the ultimate vision is for the system to operate using local models. Camel AI is now leveraging this open-source effort to attract talent, as they are rolling out industry-specific, customized multi-agent systems for enterprise clients.

### Response to Competition

- Camel AI kills Aent AI product post-Cowork announcement
- Open sources entire codebase immediately
- Renames project to openwork.ai and uses Apache 2 license
- Tweet announcing pivot garners 1.7 million views

### Multi-Agent Architecture Overview

- System built in three tiers: Electron front end
- FastAPI agent backend for orchestration and state
- Agents built on Camel AI multi-agent core
- Long-term goal is running completely with local agents

### Core Architectural Patterns

- Utilizes task decomposition hooked up to a coordinator agent
- Employs a Directed Acyclic Graph (DAG) to manage dependencies and run nodes in parallel
- Incorporates patterns from research like Magentic 1 for coordination

### Specialized Agent Capabilities

- Four pre-built agents: Developer agent (code execution, file system access)
- Browser agent (information retrieval, web search)
- Document processing agent (writing and editing)
- Multimodal agent (handling images/audio)

### Tooling and Integration

- Browser automation leverages Playwright APIs, customized into a set of specific tools for navigation and retrieval
- System demonstrated running enterprise tasks using Gemini 3 against the Salesforce platform

### Startup Strategy and Future

- Goal is hiring good agent builders for rolling out industry-specific enterprise versions
- Open sourcing aims to pay forward and advance agent use generally
- Active development and contributions continue on the GitHub repository