ContextAgent is a lightweight, context-central multi-agent systems framework designed for easy context engineering. It focuses on efficiently managing the context of each agent and binds all agents through simplified, centralized context operations. Unlike traditional multi-agent frameworks, ContextAgent treats agents simply as LLMs with different contexts, eliminating unnecessary complexity. Built with a PyTorch-like API, developers can create sophisticated multi-agent systems with minimal code.

🌟 Features

• 📋 Context = Template + State: Dynamic context management based on Anthropic's blog.
• 🔀 Decoupled Agent Design: Agent = LLM + Context. All agents are just LLMs with different contexts.
• 🎨 PyTorch-Like Pipeline API: Inherit BasePipeline, define async run(), use @autotracing for tracing.
• 🌐 Multi-LLM Support: Works with OpenAI, Claude, Gemini, DeepSeek, and more.
• 🧩 Modular Architecture: Built on OpenAI Agents SDK with clear separation: context, agents, pipeline.
• ⚡ Easy to Use & Customize: Reuse pipelines with just a query; create new ones with familiar patterns.

📢 News

• [2025-10] ContextAgent is released now!

🎬 Demo

📦 Installation

This project uses uv for fast, reliable package management.

Install uv

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

See the uv installation guide for more options.

Setup Environment

# Clone the repository
git clone https://github.com/context-machine-lab/contextagent.git
cd contextagent

# Sync dependencies
uv sync

Configure API Keys

ContextAgent requires API keys for LLM providers. Set up your environment in .env file:

# Copy the example environment file
cp .env.example .env
# Edit .env and add your API keys

See .env.example for complete configuration options.

🚀 Quick Start

Run Built-in Examples

Try out ContextAgent with pre-configured example pipelines:

Data Science Pipeline - Automated ML pipeline for data analysis and model building:

uv run python -m examples.data_science

Web Research Pipeline - Search-based research with information extraction:

uv run python -m examples.web_researcher

Basic API Pattern

Here's how to use ContextAgent in your own code:

from pipelines.data_scientist import DataScientistPipeline, DataScienceQuery

# Initialize pipeline with config
pipe = DataScientistPipeline("pipelines/configs/data_science.yaml")

# Create a query
query = DataScienceQuery(
    prompt="Analyze the dataset and build a predictive model",
    data_path="data/banana_quality.csv"
)

# Execute
pipe.run_sync(query)

Web UI (Pipeline Manager)

Run the lightweight Flask web UI to submit and monitor pipelines with live logs:

uv run python frontend/app.py --host localhost --port 9090 --debug

Then open http://localhost:9090 in your browser. The UI streams live status and panels from the running pipeline and lets you stop active runs.

Steps to Build Your Own System

🛠️ Steps to Build Your Own System

ContextAgent uses a PyTorch-like API for building multi-agent systems. Follow these steps to create your own pipeline:

Step 1 - Define Pipeline Class

Inherit from BasePipeline and call super().__init__(config):

from pipelines.base import BasePipeline
from pydantic import BaseModel

class YourPipeline(BasePipeline):
    def __init__(self, config):
        super().__init__(config)
        # Your initialization here

Step 2 - Create Context and Bind Agents

Create a centralized Context, get the LLM, and bind agents:

from contextagent.agent import ContextAgent
from contextagent.context import Context

class YourPipeline(BasePipeline):
    def __init__(self, config):
        super().__init__(config)

        self.context = Context(["profiles", "states"])
        llm = self.config.llm.main_model

        # Manager agent example
        self.routing_agent = ContextAgent(self.context, profile="routing", llm=llm)

        # Tool agents example
        self.tool_agents = {
            "data_loader": ContextAgent(self.context, profile="data_loader", llm=llm),
            "analyzer": ContextAgent(self.context, profile="analyzer", llm=llm),
            # ... add more agents
        }
        self.context.state.register_tool_agents(self.tool_agents)

Step 3 - Define Async Run with @autotracing

Define your workflow in an async run() method:

import asyncio
from pipelines.base import autotracing

class YourPipeline(BasePipeline):
    @autotracing()
    async def run(self, query: YourQuery):
        self.context.state.set_query(query)

        while self.iteration < self.max_iterations:
            self.iterate()

            # Call agents directly
            routing_result = await self.routing_agent(query)

Step 4 - Define Query Model and Execute

Create a Pydantic model and run your pipeline:

class YourQuery(BaseModel):
    prompt: str
    # Add your custom fields

# Execute
pipe = YourPipeline("pipelines/configs/your_config.yaml")
query = YourQuery(prompt="Your task here")
result = pipe.run_sync(query)

Full Example Reference

See complete implementations in:

• examples/data_science.py - Basic pipeline usage
• pipelines/data_scientist.py - Full pipeline implementation reference
• Documentation - Detailed design guide

🏗️ Architecture

ContextAgent is organized around a central conversation state and a profile-driven agent system. All agents are coordinated through a unified Context that manages iteration state and shared information.

Core Components:

• pipelines/ – Workflow orchestration and configuration management
• contextagent/agent/ – ContextAgent implementation with context awareness and execution tracking
• contextagent/context/ – Centralized conversation state and coordination
• contextagent/profiles/ – Agent profiles defining capabilities (manager, data, web, code, etc.)
• contextagent/tools/ – Tool implementations for data processing, web operations, and code execution
• examples/ – Example pipelines demonstrating usage
• frontend/ – Web UI for pipeline management and monitoring

Project Structure:

contextagent/
├── pipelines/          # Workflow orchestration
├── contextagent/
│   ├── agent/          # ContextAgent implementation
│   ├── context/        # Conversation state management
│   ├── profiles/       # Agent profiles (manager, data, web, code)
│   ├── tools/          # Tool implementations
│   └── artifacts/      # Output formatting
├── examples/           # Example pipelines
└── frontend/           # Web UI

For more details, see the full documentation.

📊 Benchmarks

ContextAgent's context-central design has been validated on multiple research benchmarks:

• Data Science Tasks: Efficient context sharing enables streamlined automated ML pipelines
• Complex Reasoning: Centralized state tracking improves multi-step reasoning coordination
• Deep Research: Search based complex reasoning and report generation

Detailed benchmark results and comparisons coming soon.

🗺️ Roadmap

• Persistence Process - Stateful agent workflows
• Experience Learning - Memory-based reasoning
• Tool Design - Dynamic tool creation
• Frontend Support - Enhanced web UI for system interaction and monitoring
• MCP Support - Full Model Context Protocol integration for extended agent capabilities
• Claude Code Skill Support - Native integration with Claude Code environment
• Workflow RAG - Retrieval-augmented generation for complex workflows

📚 Documentation

More details are available at Documentation.

🙏 Acknowledgements

ContextAgent's context-central design is inspired by the multi-agent systems research community and best practices in distributed state management. We are particularly grateful to:

• OpenAI Agents SDK - For providing a lightweight, powerful framework for multi-agent workflows and the financial research agent example that demonstrates structured research patterns.
• Youtu-Agent - For its flexible agent framework architecture with open-source model support and tool generation capabilities.
• agents-deep-research - For its iterative deep research implementation showcasing multi-agent orchestration for complex reasoning tasks.

We thank the developers of these frameworks and the broader LLM community whose work informed this architecture.

🤝 Contributing

We welcome contributions! ContextAgent is designed to be a community resource for multi-agent research. Please open an issue or submit a pull request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📖 Citation

If you use ContextAgent in your research, please cite:

@misc{contextagent2025,
  title={ContextAgent: Agent from Zero},
  author={Zhimeng Guo, Hangfan Zhang, Siyuan Xu, Huaisheng Zhu, Teng Xiao, Jingyi Chen, Minhao Cheng},
  year={2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  url={https://github.com/context-machine-lab/contextagent}
}