Awesome AI Agents › Research Platforms & Simulators

WooooDyy/AgentGym-RL

⭐ 864 Python repository created 2025-09-10

AgentGym-RL is a research framework for training large language model agents to make sequences of decisions through multi-turn interaction with real environments, rather than the single-turn math and coding settings that most RL work on LLMs targets. Its architecture is modular and separates three concerns: the environment servers, the agent that acts against them, and the training loop. Environments cover several families of realistic tasks, including web navigation, deep search, text-based games, embodied household tasks and scientific settings, and the repository documents how to plug in a custom environment. The training side extends verl and supports mainstream algorithms, so a practitioner can compare supervised fine-tuning, direct preference style methods and policy-gradient RL under one interface. The paper accompanying the code proposes ScalingInter-RL, a schedule that progressively lengthens the allowed agent-environment interaction horizon during training: short horizons early keep variance manageable and encourage exploitation of basic skills, and the horizon is extended as the policy stabilises so the agent learns longer-range behaviour without the optimisation collapsing. The authors report that a 7B open model trained this way reaches or exceeds commercial model performance on 27 tasks spanning the included environments. A visual interactive interface is provided for replaying and examining complete interaction trajectories, which makes it possible to inspect where an episode went wrong instead of reading raw logs. The RL dataset and benchmark are published on Hugging Face as AgentGym-RL-Data-ID, and the paper was accepted at ICLR 2026 as an oral. The project is written in Python and is aimed at researchers and engineers doing agent post-training, environment design or reproducible evaluation of long-horizon agent behaviour.

https://github.com/WooooDyy/AgentGym-RL

reinforcement-learningllm-agentsmulti-turntraining-frameworkbenchmarkenvironmentsverl