AgentGym: 14 Interactive Environments Behind One ReAct Interface

The Fragmentation Problem AgentGym Was Built To Fix

Before a framework like this exists, an agent researcher who wants to test a model on WebShop, ALFWorld and a text game writes three adapters. Each original repository defines its own observation string, its own action grammar, its own episode termination rule and its own scoring script. The comparison you then publish is partly a comparison of your adapters. AgentGym's stated goal is to remove that layer: the introduction describes it as featuring "a variety of environments and tasks for broad, real-time, uniformat, and concurrent agent exploration." The word uniformat is doing the work. Fourteen environments are listed in the README table, spanning web navigation, text games, household tasks, digital games, embodied tasks, tool use and programming, and each is mapped to an EnvServer directory in the repository. The intended audience is narrow and identifiable: groups doing agent training and evaluation research who need cross-environment numbers without maintaining fourteen forks. If you only ever evaluate on one environment, the abstraction costs you more than it returns.

One ReAct Protocol, Fourteen Environment Servers

The architecture visible in the repository layout is a client-server split. Each environment has its own agentenv-* directory: agentenv-webshop, agentenv-webarena, agentenv-lmrlgym (which serves both MAZE and Wordle), agentenv-alfworld, agentenv-sciworld, agentenv-babyai, agentenv-textcraft, agentenv-tool (Weather, Movie, Academia, Sheet and TODOList), and agentenv-sqlgym for BIRD. The agent side speaks the ReAct format, and the environment side is a server process. That split is what makes concurrency possible: the README notes an update that improved "support for large-scale parallel execution (e.g., parallel runs in WebArena)." It is also the main operational cost, because you are not installing a library, you are deploying a set of services. The environments are not native to AgentGym. The table lists an Original Repo for each one (WebShop-Repo, WebArena, LMRL-Gym, ALFWorld-Repo, ScienceWorld, babyai, ADaPT for TextCraft, AgentBoard for the four tool environments, and AlibabaResearch/DAMO-ConvAI for BIRD), which tells you the EnvServer is a wrapper around someone else's simulator. When that upstream simulator needs a database, a browser stack or a specific Python version, AgentGym inherits the requirement.

What Ships Alongside The Code: AgentTraj, AgentEval, AgentEvol-7B

Three artifacts are published separately from the platform code, and they are the reason the repository is useful even to people who never run an environment server. AgentTraj-L is a trajectory dataset on Hugging Face. AgentEval is a benchmark suite. AgentEvol-7B is a released model. The trajectory counts in the README table vary enormously by environment: WebShop has 3930 trajectories and 200 evaluation tasks, ALFWorld 2420 and 200, SciWorld 2120 and 200, BIRD 3000 and 200, while WebArena, Academia and Sheet have 0 trajectories and 20 evaluation tasks each. That asymmetry is worth reading carefully. The environments with zero trajectories are the ones where collecting interaction data is hardest or most expensive, so they contribute evaluation only. If your work depends on imitation learning from trajectories, the usable subset is smaller than the fourteen-environment headline suggests. AgentEvol is the method described in the original paper: the README states the goal was to "investigate the potential of agent self-evolution beyond previously seen data across tasks and environments," and reports that evolved agents reach results comparable to SOTA models. That claim belongs to the paper, not to anything a reader can verify from the repository alone.

Getting It Running: Servers First, Agent Second

The README does not inline an installation block in the material available here, so treat the repository's own documentation as the source of truth rather than any command reconstructed from memory. What the material does establish is the shape of the setup. You clone the repository, and for each environment you intend to use you start the corresponding agentenv-* server, then point your agent at it. The custom-environment path is documented explicitly: the 2025-09-10 news entry says "You can develop your custom environment to AgentGym and perform RL on it! The tutorial is here," linking to docs/tutorials/en/05-2nd-Development.md. That file is the one to read before writing any integration code, because it defines the contract an EnvServer must satisfy. Two operational notes matter. First, the 2025-09-03 entry mentions an interactive frontend for visualization that lets researchers "replay and inspect full trajectories, step through agent decision-making," which is the practical debugging surface when an episode fails. Second, the same entry says several environments were updated for stability and parallel execution, with WebArena called out by name. WebArena is the environment most likely to dominate your setup time, since it is a browser-based benchmark and the README's own table shows it contributing 20 evaluation tasks and no trajectories.

The Uncomfortable Parts: Environment Weight And A Moving Target

The clearest limitation is that AgentGym does not abstract away the environments' dependencies, only their interfaces. Fourteen servers means fourteen sets of Python packages, and for WebArena that includes a real browser automation stack. Reproducing a full cross-environment sweep is a systems project before it is a research project. A second issue is the zero-trajectory environments. WebArena, Academia and Sheet appear in the table with 0 trajectories and 20 eval tasks, so any claim about training across all fourteen environments is really a claim about training across eleven. Third, the repository has moved on. On 2025-09-10 the maintainers announced AgentGym-RL, described as "the reinforcement learning (RL) version of AgentGym, enabling agents to learn directly from interactive environments," alongside a new arXiv paper on multi-turn RL for long-horizon decision making. The original AgentGym remains the supervised and evaluation harness; the RL work lives in a separate repository. Anyone arriving now should decide which of the two they actually want before investing in environment setup, because the environment servers are shared but the training code is not. Finally, the README's own update log, with three entries dated 2025-09-03 and 2025-09-10, signals active churn in the environment wrappers. Pin a commit if you need a stable baseline for a paper.

How AgentGym Differs From AgentBoard

The most direct comparison available from the material is AgentBoard, because AgentGym's own table lists AgentBoard repositories as the origin of the Weather, Movie, Academia, Sheet and TODOList environments, all served through agentenv-tool. AgentBoard is an evaluation benchmark: it defines tasks and metrics for assessing an agent's multi-turn ability. AgentGym takes those environments and adds a training and evolution layer around them, plus a trajectory dataset and a released model. The difference in approach is scope rather than quality. If your question is "how good is my model on these five tool-use tasks," AgentBoard answers it directly and you avoid the AgentGym server layer entirely. If your question is "can a model improve across web, embodied and tool environments simultaneously," you need the uniform action format and the trajectory corpus, which is what AgentGym provides. Choosing AgentGym for pure evaluation means paying for infrastructure you will not use.

Licence, Maintenance And What To Verify Before You Commit

The repository is MIT licensed, which is permissive and imposes few obligations on your own code. The complication is that AgentGym wraps upstream projects, and those carry their own licences. WebShop, WebArena, ALFWorld, ScienceWorld, BabyAI, ADaPT and the DAMO-ConvAI BIRD code are separate works with separate terms, and the README's table links to each. MIT on the AgentGym repository does not relicense them, and it says nothing about the Hugging Face datasets AgentTraj-L and AgentEval or the AgentEvol-7B weights, which are distributed on Hugging Face under whatever terms those model and dataset cards specify. Check each one for the environments you plan to use. This is not legal advice; if you are shipping something commercial, have someone read the upstream licences. On maintenance, the material supports a mixed read: the last push is dated 2026-05-30 and the news log shows environment stability fixes in September 2025, so the project is alive, but there are no retrieved releases, which means you are tracking a branch rather than versioned artifacts. The concrete next step is to read docs/tutorials/en/05-2nd-Development.md, count how many of the fourteen agentenv-* servers you can actually host in your own infrastructure, and pin a commit hash before running anything you intend to publish.

Editorial conclusion

Adopt AgentGym if you are reproducing the AgentEvol experiments or need several heterogeneous agent benchmarks under one action format, and accept that each agentenv-* server is a separate service you must stand up. Do not adopt it if you want a single install that evaluates a production agent, because the suite assumes you can host WebShop, WebArena, ALFWorld and the rest yourself. Before committing, read docs/tutorials/en/05-2nd-Development.md and confirm which of the fourteen environments you can actually host, since the WebArena and SQLGym servers carry the heaviest external dependencies.

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. WooooDyy/AgentGym on GitHub