English | 中文
Phoenix Agent is a general-purpose AI Agent system that supports safe execution of various tools and operations in a sandboxed environment.
Phoenix Agent is a Hierarchical Multi-Flow Agent System that moves beyond generic single-flow architectures by implementing task-specific workflows. Our system achieves 73.6% accuracy on GAIA Level 1 tasks through specialized flows optimized for distinct task domains.
Phoenix Agent implements a hierarchical architecture where a global Orchestrator coordinates specialized sub-flows:
- Search Flow: Recursive "Knowledge Gap" decomposition for application-oriented search
- Code Flow: Dual-layer memory system with remote sandbox for isolated execution
- Analysis Flow: Specialized reasoning workflows for complex problem-solving
- Hierarchical Multi-Flow Design: Task-specific workflows instead of generic pipelines
- Intelligent Orchestration: Central coordinator routes subtasks to specialized flows
- Dual-Layer Memory System: "Cold and Hot" memory for efficient context management
- Secure Sandbox Environment: Isolated execution environment based on Docker/Kubernetes
- Rich Tool Integration: Browser automation, Shell execution, file operations, search engines, etc.
- Real-time Communication: SSE-based streaming event transmission
- Visualization: VNC remote desktop support
- Flexible Deployment: Supports Docker Compose and Kubernetes/Helm deployment
📄 Technical Report: For detailed architecture, methodology, and experimental results, see our technical report:
Beyond Generic Agents: A Hierarchical Multi-Flow Agent System with Task-Specific Workflows
Figure 1: Overview of the Hierarchical Multi-Flow Architecture with Task-Specific Workflows
When a user initiates a conversation:
- Web sends a request to create an Agent to the Server, which creates a Sandbox through
/var/run/docker.sockand returns a session ID. - The Sandbox is an Ubuntu Docker environment that starts Chrome browser and API services for tools like File/Shell.
- Web sends user messages to the session ID, and when the Server receives user messages, it forwards them to the PlanAct Agent for processing.
- During processing, the PlanAct Agent calls relevant tools to complete tasks.
- All events generated during Agent processing are sent back to Web via SSE.
When users browse tools:
- Browser:
- The Sandbox's headless browser starts a VNC service through xvfb and x11vnc, and converts VNC to websocket through websockify.
- Web's NoVNC component connects to the Sandbox through the Server's Websocket Forward, enabling browser viewing.
- Other tools: Other tools work on similar principles.
Phoenix Agent has been evaluated on the General AI Assistants (GAIA) benchmark, demonstrating strong performance across multiple AI models. Our hierarchical multi-flow architecture achieves:
| Metric | Score |
|---|---|
| Level 1 | 73.6% |
| Level 2 | 62.37% |
| Level 3 | 34.59% |
| Overall | 22.45% |
These results indicate that a hierarchical multi-flow design with task-specific workflows can achieve non-trivial performance on GAIA-style tasks without task-specific fine-tuning, demonstrating the advantage of moving beyond generic agent architectures.
Phoenix Agent supports and has been tested with the following AI models:
- GLM 4.5
- GLM 4.5 Air
- Gemini 2.5 Pro
- GPT 4.1
- Sense Voice Small
For detailed experimental setup, methodology, and analysis, please refer to our technical report.
- Technical Report - Complete technical report on hierarchical multi-flow architecture, methodology, and experimental results
- Architecture Design - Detailed system architecture and workflow design
- Deployment Guide - Step-by-step deployment instructions for production and development environments
- Environment Variables Guide - Complete reference for all configuration options
Phoenix Agent is developed by:
- Yufeng Lin
- Yuzhong Zhang
- Liwei Liu
- Yimeng Teng
- Wentao Lin
- Yao Li
- Ming Wen
- Xuhuan Shen
Contact: {yufenglin, yuzhongzhang, liweiliu, yimengteng, wentaolin, yaoli, mingwen, xuhuanshen}@link.cuhk.edu.cn
- ai-manus - Original project inspiration