Youtu-Agent

What it does

• Two generation paradigms: Workflow mode for standard tasks and a Meta-Agent mode that interviews you, assembles tools, and writes YAML configs — the README reports an 81%+ tool synthesis success rate
• Benchmark results on open-weight models: GAIA 72.8% pass@1 with DeepSeek-V3 (text-only validation subset) and WebWalkerQA 71.47% with DeepSeek-V3.1
• Agent Practice module: Training-Free GRPO accumulates experience in-context without parameter updates (e.g. +5.4% on AIME 2025, ~$8 per RL run)
• Agent RL module: end-to-end RL pipeline integrated with distributed frameworks — 40% training speedup, scaling to 128 GPUs
• Agent Skills supported since 2026-01: extend agents with modular, domain-specific workflows inspired by Claude Code skills; built on openai-agents, runs fully asynchronously

When to use it

Three audiences map to the README's own split: researchers get a strong open baseline with one-click evaluation scripts (WebWalkerQA, GAIA); application developers get portable scaffolding with YAML configs and built-in toolkits; enthusiasts get ready-made examples — deep research reports, CSV data analysis, file organization, RAG, and PPT generation.

Getting started

Requires Python 3.12+ and uv:

git clone https://github.com/TencentCloudADP/youtu-agent.git
cd youtu-agent
uv sync            # or: make sync
source ./.venv/bin/activate
cp .env.example .env   # then fill in LLM keys (OpenAI-compatible format)

Run a first agent:

python scripts/cli_chat.py --config simple/base_search

Web search additionally needs SERPER_API_KEY and JINA_API_KEY in .env. Docker-based setup with a web UI: see docker/README.md in the repo.

Things to know

Benchmark numbers are vendor-run on specific subsets (GAIA figure is the text-only validation subset) — treat them as a baseline to reproduce, not a guarantee. The search toolkit depends on third-party API keys (Serper, Jina). Docs live on a separate site (tencentcloudadp.github.io/youtu-agent), and the RL/Agent-Lightning integration sits on a dedicated branch rather than main.