Youtu-Agent Project Details

Project Overview

Youtu-Agent is a flexible, high-performance agent framework specifically designed for building, running, and evaluating autonomous intelligent agents. This framework not only demonstrates excellent performance in benchmarks but also provides powerful agent capabilities based on open-source models, including features such as data analysis, file processing, and in-depth research.

Project Origin

• Developed by: Tencent Youtu Lab
• Project Repository: https://github.com/TencentCloudADP/youtu-agent
• Open-Source License: Fully open-source
• Base Framework: Built upon openai-agents

Core Features

🌟 Verified Performance

• WebWalkerQA: Achieved 71.47% accuracy (pass@1) using DeepSeek-V3.1
• GAIA Benchmark: Scored 72.8% pass@1 on the text subset
• Purely Open-Source Models: Fully utilizes DeepSeek-V3 series models, without reliance on Claude or GPT

🚀 Open-Source Friendly and Cost-Conscious

• Optimized for accessible, low-cost deployment, eliminating the need for closed-source models
• Supports various model APIs, from DeepSeek to gpt-oss
• Features extensible tool integration and framework implementation

💡 Practical Application Scenarios

The framework provides out-of-the-box task support for:

• CSV data analysis
• Literature reviews
• Personal file organization
• Podcast and video generation (coming soon)
• In-depth research reports
• Web page generation

🛠️ Flexible Architecture Design

• Modular Design: Key components like Environment and ContextManager are well-encapsulated and highly customizable
• YAML Configuration: Structured configuration based on YAML, easy to manage
• Fully Asynchronous: Achieves high performance and efficient execution, especially suitable for benchmark evaluation
• Tracing and Analysis: Provides in-depth analysis of tool calls and agent trajectories

System Architecture

Core Components

1. Agent: An LLM configured with specific prompts, tools, and an environment
2. Toolkit: A collection of encapsulated tools that an agent can use
3. Environment: The world in which an agent operates (e.g., browser, Shell)
4. ContextManager: A configurable module that manages the agent's context window
5. Benchmark: An encapsulated workflow for specific datasets

Agent Modes

The framework supports two primary agent modes:

• SimpleAgent: A classic single-agent model employing a reasoning-action loop (ReAct style)
• OrchestraAgent: A multi-agent system using a plan-execute strategy, coordinating a planner, workers, and a reporter

Installation and Quick Start

System Requirements

• Python 3.12+
• uv recommended for dependency management

Installation Steps

# Clone the repository
git clone https://github.com/TencentCloudADP/youtu-agent.git
cd youtu-agent

# Sync dependencies
uv sync # Or use make sync

# Activate virtual environment
source ./.venv/bin/activate

# Copy environment configuration file
cp .env.example .env

Environment Configuration

Configure necessary API keys in the .env file:

# LLM Configuration (DeepSeek Example)
UTU_LLM_TYPE=chat.completions
UTU_LLM_MODEL=deepseek-chat
UTU_LLM_BASE_URL=https://api.deepseek.com/v1
UTU_LLM_API_KEY=replace-to-your-api-key

# Tencent Cloud DeepSeek API (New users can get 3 million free tokens)
UTU_LLM_TYPE=chat.completions
UTU_LLM_MODEL=deepseek-v3
UTU_LLM_BASE_URL=https://api.lkeap.cloud.tencent.com/v1
UTU_LLM_API_KEY=replace-with-your-api-key

# Tool API Keys
SERPER_API_KEY=<Get from https://serper.dev/playground>
JINA_API_KEY=<Get from https://jina.ai/reader>

Usage

1. Basic Chat

# Launch an interactive chatbot with default configuration
python scripts/cli_chat.py --stream --config default

# Basic version without search tools
python scripts/cli_chat.py --stream --config base

2. Automated Agent Configuration Generation

A prominent feature of Youtu-Agent is its ability to automatically generate agent configurations. Unlike other frameworks that require writing code or meticulously crafting prompts, Youtu-Agent uses simple YAML configurations, with a built-in "meta-agent" that converses with you to understand requirements, then automatically generates and saves the configuration.

# Interactively generate configuration
python scripts/gen_simple_agent.py

# Run the generated configuration
python scripts/cli_chat.py --stream --config generated/xxx

3. Example Applications

SVG Image Generation

python examples/svg_generator/main.py

Web UI Version

# Download the frontend package
curl -LO https://github.com/Tencent/Youtu-agent/releases/download/frontend%2Fv0.1.5/utu_agent_ui-0.1.5-py3-none-any.whl

# Install the frontend package
uv pip install utu_agent_ui-0.1.5-py3-none-any.whl

# Run the Web version
python examples/svg_generator/main_web.py

Access http://127.0.0.1:8848/ to use the Web interface.

4. Benchmark Evaluation

# Prepare dataset (download and process WebWalkerQA dataset)
python scripts/data/process_web_walker_qa.py

# Run evaluation
python scripts/run_eval.py --config_name ww --exp_id <your_exp_id> --dataset WebWalkerQA_15 --concurrency 5

Real-world Use Cases

Data Analysis

Agents can analyze CSV files and generate HTML reports, automatically identifying data patterns, generating visualizations, and insightful reports.

File Management

Agents can rename and categorize local files, automatically organizing document structures based on content.

Extensive Research

Similar to Manus functionality, collecting vast amounts of information to generate comprehensive research reports.

Paper Analysis

Parsing given papers, performing analysis, and compiling relevant literature to produce final results.

Technical Highlights

1. Minimalist Design

The framework strives to remain simple and easy to use, avoiding unnecessary overhead.

2. Streaming and Tracing

Based on the openai-agents SDK, it inherits streaming, tracing, and agent loop capabilities.

3. In-depth Analysis System

In addition to OTEL, the DBTracingProcessor system provides in-depth analysis of tool calls and agent trajectories.

4. Trajectory Collection

Supports data collection for training and research purposes.

Target User Groups

Researchers

• Provides powerful open-source benchmarks and a flexible architecture
• Supports model training, evaluation, and ablation studies
• One-click evaluation scripts simplify experimental processes

Application Developers

• Practical use cases and user-friendly scaffolding
• Rapidly build real-world agent applications

AI and Agent Enthusiasts

• Rich examples and intuitive development tools
• Easy to get started, suitable for learning and exploration

Enterprise Tech Teams

• Supports low-cost deployment and open-source models
• Applicable for intelligent customer service, data analysis, and other applications in enterprise environments

Educators and Students

• Offers features like personalized learning and automated grading
• Aids in teaching and learning processes

Future Development

The project roadmap includes:

• Tool evaluation and automated optimization
• Custom tool generation
• Expanded multimodal tool integration
• Enhanced benchmark coverage
• Refined tracing and analysis system

Community and Support

• Documentation: https://tencentcloudadp.github.io/youtu-agent/
• Discord Community: https://discord.gg/svwuqgUx
• Enterprise Solutions: Tencent Cloud Agent Development Platform (ADP)

Citation

If you find this project useful, please consider citing:

@misc{youtu-agent-2025,
title={Youtu-agent: A Simple yet Powerful Agent Framework},
author={Tencent Youtu Lab},
year={2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/TencentCloudADP/youtu-agent}},
}

Conclusion

Youtu-Agent is a powerful and easy-to-use open-source agent framework, providing developers with a complete solution for building high-performance AI applications. Its design philosophy based on open-source models, excellent performance, and rich feature set make it an important tool in the field of agent development. Whether for research or commercial applications, Youtu-Agent can provide users with a reliable and efficient agent solution.