Youtu-Agent

Youtu-Agent is a modular Python framework from Tencent (TencentCloudADP) for building, running, and evaluating autonomous agents. It is built on the openai-agents SDK and driven by YAML configuration, with pre-built toolkits, environments, and an evaluation pipeline. The latest listed release is v0.1.3.

What It Is

Youtu-Agent is an agent framework in which an AgentConfig defines an Agent, the Agent operates within an Environment and uses Toolkits to act, and an Evaluation Framework benchmarks the result. Configuration is built on pydantic and hydra. The code abbreviates the project as utu. It is aimed at agent researchers, LLM trainers, and application developers.

Agent Paradigms and Toolkits

Two paradigms are supported: SimpleAgent, a single-agent ReAct-style loop, and OrchestraAgent, a multi-agent Plan-and-Execute setup coordinating a Planner, Workers, and a Reporter. Environments include a local shell environment for filesystem access and a browser environment for web interaction. Pre-built toolkits cover web search, file manipulation, code execution, document analysis, image, video, bash, and Python execution. Agent Skills, which are modular domain-specific knowledge and workflows, were added in January 2026.

Automated Generation and Learning

A built-in meta-agent can interview the user about requirements, then generate tool code, prompts, and YAML configs that can be run with a CLI chat script. The Agent Practice module uses Training-Free GRPO to improve agents through in-context experience without parameter updates. An Agent RL pipeline integrates with distributed training frameworks. The README attributes benchmark results, such as 71.47% on WebWalkerQA and 72.8% on the GAIA text-only validation subset, to the project's own evaluations.

Setup Path

The project requires Python 3.12+ and recommends uv. Users clone the repository, sync dependencies, and copy .env.example to .env with an OpenAI-API-compatible LLM key. Examples include data analysis, file organization, deep and wide research, paper analysis, RAG, and PPT generation. Web search examples need Serper and Jina API keys. A Docker setup and a web UI are also documented, and evaluation scripts support benchmarks such as WebWalkerQA.