Youtu-Agent is a flexible, high-performance agent framework for building, running, and evaluating autonomous agents. Built on open-source models, it can provide capabilities such as data analysis, file processing, and deep research.
Key Capabilities
- Proven performance: When using only open-source weight models (such as DeepSeek-V3), it achieves 71.47% accuracy on WebWalkerQA and 72.8% pass@1 on the GAIA text validation subset, establishing a strong open-source baseline.
- Automated agent generation: It provides two paradigms, Workflow mode (standard tasks) and Meta-Agent mode (complex requirements), to automatically generate tool code, prompts, and configurations, with a tool synthesis success rate exceeding 81%.
- Continuous experience learning: The Agent Practice module enables low-cost continuous evolution through Training-Free GRPO, without parameter updates, improving by 5.4% on AIME 2025.
- Scalable and stable agent reinforcement learning: The Agent RL module provides a complete end-to-end RL pipeline, integrates with distributed frameworks, accelerates training by 40%, and can scale to 128 GPUs.
- Open-source friendly and cost-controllable: Optimized for low-cost deployment and does not depend on closed-source models.
- Practical use cases: Out-of-the-box support for data analysis, literature review, personal file organization, retrieval-augmented generation (RAG), and PPT generation, among others.
- Flexible architecture: Built on openai-agents, extensible to support multiple model APIs (from DeepSeek to gpt-oss), tool integrations, and framework implementations.
Usage
The project requires Python 3.12+ and recommends using uv to manage dependencies. Basic steps: clone the repository and sync dependencies, copy .env.example and fill in the required keys (such as an LLM API key), then run the built-in configuration to start an interactive CLI chatbot. For example, configs/agents/simple/base_search.yaml defines a simple agent with a search tool.
Some examples require web search capability, and tool APIs need to be configured in the tools module of .env. The repository provides multiple ready-to-use examples, such as having the agent automatically search for information and generate an SVG image on the theme "DeepSeek V3.1 New Features"; a Web UI for visualizing runtime status can also be obtained by downloading the frontend package. It supports running on standard datasetsfor evaluation, such as WebWalkerQA, and the results can be stored and analyzed on the evaluation platform.
Applicable Scenarios
- Agent researchers and LLM trainers: Provides a simple and powerful baseline and one-click evaluation scripts, facilitating model training and ablation studies.
- Agent application developers: A portable scaffold, simple scripts, and a rich built-in toolkit, with a modular design (Environment, ContextManager, etc. are customizable).
- AI and agent enthusiasts: The /examples directory contains tasks such as deep research report generation, data analysis, and personal file organization. The rich toolset and visual tracing facilitate development and debugging.
Limitations and Notes
The documentation does not provide information on fees, pricing, or commercial licensing. Some features (such as the DBTracingProcessor analysis system) are marked as "coming soon." The framework depends on open-source models and external tool APIs, and specific availability is limited by the configured models and keys.
