AgentGym

Overview of AgentGym

What it is

AgentGym is a framework for evaluating and developing large language model (LLM) based agents across diverse environments. It is the code and implementations for the ACL 2025 paper by Zhiheng Xi et al. The framework includes a platform with 14 interactive environments, a high-quality trajectory set (AgentTraj-L), a benchmark suite (AgentEval), and a method for agent self-evolution (AgentEvol). It also provides a reinforcement learning version (AgentGym-RL) and an interactive frontend for visualization.

Key features

• Supports 14 environments across web navigating, text games, house-holding tasks, digital games, embodied tasks, tool-using, and programming.
• Uses a unified ReAct format for agent-environment interaction.
• Provides real-time feedback and supports concurrent execution.
• Includes the AgentTraj-L trajectory set and the AgentEval benchmark on Hugging Face.
• Introduces the AgentEvol method for evolving agents beyond previously seen data.
• Offers an RL version (AgentGym-RL) for multi-turn reinforcement learning.
• Allows developers to easily add custom environments via encapsulated HTTP services.

Use cases

• Evaluating and developing generally-capable LLM-based agents.
• Researching agent self-evolution across tasks and environments.
• Training agents through multi-turn reinforcement learning.
• Replaying and inspecting full agent trajectories for analysis.

Getting started

See the repository README for installation and quick start instructions.

License and pricing

The project is licensed under MIT and is open source.

Based on the project's public README and repository metadata.