A gym where the treadmill is a web browser

What it does BrowserGym turns web automation into a Gymnasium environment: your agent gets observations, returns actions, and receives rewards, exactly like any other RL setup — except the environment is a real Chromium session driven by Playwright. It ships with nine benchmarks baked in, including MiniWoB, WebArena, VisualWebArena, WorkArena, AssistantBench, WebLINX, OpenApps, and TimeWarp. There’s also an openended task that just drops your agent on any URL and lets it chat with a user.

The interesting bit The unglamorous part is the point: every benchmark speaks the same env.step() dialect, so you can swap tasks without rewriting your agent loop. Building your own benchmark means inheriting one class, AbstractBrowserTask — the framework is deliberately extensible rather than a fixed test suite.

Key highlights

• Standard Gymnasium API (gym.make, reset, step) over real browser sessions
• Nine bundled benchmarks, from synthetic MiniWoB pages to realistic WorkArena tasks
• Open-ended interactive mode where the agent converses with a user on any start URL
• Companion project AgentLab handles running, tracing, and evaluating agents across benchmarks
• Published in TMLR (2025), with a demo agent included (OpenAI-backed, gpt-4o-mini by default)

Caveats

• Explicitly a research framework, not a consumer product — the README says so in a warning box
• Each benchmark carries its own additional setup beyond the base install; some (WebArena, WorkArena) need self-hosted infrastructure
• The demo agent requires an OpenAI API key

Verdict If you build or evaluate web agents, this is close to the de facto standard harness — the ecosystem (AgentLab, DoomArena) is built around it. If you just want a browser to script, plain Playwright is lighter.