Training and evaluating modern AI agents with modular tasks, tools, and backends.
- 🚀 Initial public release with SWE, MemAgent (step-wise training), and Web Research examples!
- Unified interface for agentic tasks and training backends
- Pluggable tools (browser, search, code execution, finish, etc.)
- Efficient and flexible async dispatching strategies
- Works with OpenAI-compatible serving (vLLM/others), VERL, SkyRL-Train, Tinker. Switch the backend with one line configuration!
git clone --recurse-submodules https://github.com/NovaSky-AI/SkyRL.git
# our working directory
cd skyrl-agentThen head to the examples/ folder to run tasks (training and inference). Each task’s script/YAML documents its own knobs and environment requirements.
- OSWorld Integration
- Simplify SWE agent training code path
- More training recipes
- Evaluation harness unification
Huge thanks to these projects:
@article{cao2025skyrl,
title={SkyRL-Agent: Efficient RL Training for Multi-turn LLM Agent},
author={Cao, Shiyi and Li, Dacheng and Zhao, Fangzhou and Yuan, Shuo and Hegde, Sumanth R and Chen, Connor and Ruan, Charlie and Griggs, Tyler and Liu, Shu and Tang, Eric and others},
journal={arXiv preprint arXiv:2511.16108},
year={2025}
}