AgentGym: An Open-Source Framework for Self-Evolution of Large Language Model Agents in Diverse Environments

AgentGym is an open-source framework developed by research teams from institutions such as Fudan University and Alibaba. It supports training, evaluating, and evolving large language model (LLM)-based agents in 14 different environments. The framework includes three core components: the AgentTraj trajectory dataset, AgentEval benchmark, and AgentEvol evolution algorithm. It aims to promote the development of general-purpose LLM agents, lower the research threshold in this field, and provide a unified platform for academia and industry.

Traditional agent research is often limited to single environments or specific tasks, making it difficult to evaluate general capabilities. Existing benchmarks mostly focus on static dataset performance and lack systematic evaluation in dynamic interactive environments. The design philosophy of AgentGym is: a truly general-purpose agent should be able to interact, learn, and evolve in real time in diverse environments, and needs to possess multiple capabilities such as language understanding, reasoning, planning, tool use, and environmental adaptation.

1. AgentGym Platform: Provides 14 diverse interactive environments (covering categories such as web navigation, text games, household tasks, tool use, programming, etc.), uniformly uses ReAct format for interaction, supports real-time feedback and concurrent execution, and allows expansion of new environments.
2. AgentTraj-L Trajectory Dataset: Contains interaction trajectories in 14 environments, recording the agent's thinking, decision-making, and environmental feedback, covering both successful and failed cases of tasks at different difficulty levels.
3. AgentEval Benchmark: A comprehensive evaluation suite covering 14 environments, establishing unified evaluation standards, including hard indicators such as task completion rate and soft indicators such as reasoning quality and action efficiency.

AgentEvol Evolution Algorithm: Enables agents to learn from trial and error in multiple environments, accumulate cross-environment general skills, and form robust behavior patterns. Its experimental performance reaches the current advanced level. AgentGym-RL Extension: To be released in September 2025, it introduces reinforcement learning, supports training for long-cycle decision-making tasks, optimizes the agent's long-term decision-making ability, and achieves a leap from supervised learning to reinforcement learning.

1. Standardized Evaluation: Provides a fair comparison benchmark for different teams, promoting overall progress in the field;
2. Rapid Prototype Development: Uses existing environments and datasets to quickly verify new architectures and training methods;
3. Cross-Environment Transfer Learning: Diverse environments help research the transfer capabilities of agents and explore the path to general AI;
4. Community Collaboration: The open-source nature encourages global researchers to contribute new environments and improvements, forming a healthy ecosystem.

The team will continue to expand the coverage of environments, optimize training efficiency, and explore more advanced evolution algorithms. AgentGym lowers the entry threshold in the LLM agent field and provides an important platform for the development of general-purpose AI agents. With community contributions and technological iterations, we look forward to more innovative agent applications based on AgentGym.