OxyGent is a Python framework for building and deploying production-ready multi-agent AI systems. It targets developers and enterprises seeking to create efficient, scalable, and adaptable AI teams with transparent, end-to-end pipelines, enabling seamless collaboration and continuous evolution of intelligent agents.

How It Works

OxyGent employs a modular, component-based architecture where agents, tools, and LLMs are standardized "Oxy" components that snap together. This design facilitates rapid assembly, hot-swapping, and cross-scenario reuse of agents through clean Python interfaces, avoiding complex configurations. Its elastic architecture supports various agent topologies, from simple ReAct to complex hybrid planning, with automated dependency mapping and visual debugging for performance optimization.

Quick Start & Requirements

• Installation: pip install oxygent or uv pip install oxygent. Development setup requires Node.js and pip install -r requirements.txt.
• Prerequisites: Python 3.10+, conda or uv for environment management. LLM API keys and base URLs are required for operation.
• Setup: Basic setup involves environment creation and package installation, estimated to be under 10 minutes.
• Documentation: Document

Highlighted Details

• Achieves 59.14 points on the GAIA benchmark, close to the top open-source system OWL (60.8 points).
• Enables agents to dynamically plan, decompose tasks, and negotiate solutions with real-time adaptation.
• Supports a complete production lifecycle: code, deploy, monitor, and evolve agents.
• Offers full transparency and auditability of agent decisions.

Maintenance & Community

The project is actively maintained by JD.com. Community support is available via the project's GitHub Issues.

Licensing & Compatibility

• License: Apache License 2.0.
• Compatibility: Permissive license suitable for commercial use and integration with closed-source systems.

Limitations & Caveats

The framework is designed for Python 3.10 and may require specific environment configurations. While performance is benchmarked, real-world scalability and efficiency will depend on the specific agent implementations and underlying infrastructure.