youtu-agent

Why we included this project

Teams that want to run autonomous agents on open-weight models instead of paying for closed APIs get a framework that has already put up strong numbers on GAIA and WebWalkerQA using DeepSeek-V3, so the performance story has evidence behind it. The runtime covers the usual agent work, data analysis, file processing, and deep research, out of the box. Two things set it apart: an automated generation path that writes the tool code, prompts, and configuration for a described capability, and a practice module that sharpens performance on specific tasks through training-free GRPO rather than fine-tuning. When training is the goal, the full RL pipeline scales to multi-GPU setups. It builds on the openai-agents base, which keeps it familiar to Python developers and easy to point at different model providers.