youtu-agent

Youtu-Agent is a flexible framework for building, running and evaluating autonomous agents, developed by Tencent Cloud ADP and released as open source. It is built on top of the openai-agents SDK, inheriting streaming, tracing and agent-loop capabilities, and is designed to work with open-weight models such as DeepSeek-V3 rather than depending on closed models. Core concepts include the Agent (an LLM with prompts, tools and an environment), Toolkit (a packaged set of tools), Environment (the world the agent operates in, such as a browser or shell), ContextManager (context-window management) and Benchmark (an encapsulated dataset workflow with preprocessing, rollout and judging). Notable capabilities described in the README: - Automated agent and tool generation: a Workflow mode for standard tasks and a Meta-Agent mode for complex requirements. The meta-agent can interview the user, assemble tools, generate YAML configurations and produce tool code and schemas. The README reports over 81% tool synthesis success rate. - Continuous experience learning: an Agent Practice module based on Training-Free GRPO, which keeps the model frozen and learns from in-context experience without parameter updates. - Agent RL: a pipeline for end-to-end reinforcement learning, integrated with distributed frameworks such as Agent-Lightning, described as scaling to 128 GPUs. - Benchmarking: the project reports results on WebWalkerQA and GAIA using open-weight models, and ships evaluation scripts and an analysis frontend. - Practical examples: data analysis, file management, wide/deep research, paper analysis, RAG integration with RAGFlow, SVG generation and PPT generation. Getting started requires Python 3.12+ and uv for dependency management. Users clone the repository, run uv sync, copy .env.example to .env and configure LLM API keys (the README shows DeepSeek-compatible endpoints). A Docker-based setup with an interactive frontend is also documented. Agents are defined through YAML configs, and an interactive CLI chatbot can be launched with scripts/cli_chat.py. Web search examples require SERPER_API_KEY and JINA_API_KEY. A web UI package can be installed from releases to visualize agent runtime. The framework is fully asynchronous, supports both responses and chat.completions APIs, and includes a DBTracingProcessor system for analyzing tool calls and trajectories. The README positions the project for agent researchers and LLM trainers seeking a baseline, application developers needing portable scaffolding, and enthusiasts wanting practical, debuggable examples.