WebArena

What does WebArena do?

WebArena runs a set of realistic web environments and verifiable tasks that let browser agents practice long-horizon web work through the front end. It combines generated websites, annotated programs, and programmatic verifiers so tasks can be checked for functional correctness instead of judged by hand. The project spans multiple benchmark variants, including WebArena, WebArena-Infinity, VisualWebArena, and TheAgentCompany, and the environments can embed tools and knowledge resources as separate websites to mirror real workflows. At scale, the Infinity pipeline can generate each environment within ten hours at a cost under $100, and the release includes ten environments, 1,260 verifiable tasks, and 2,070 successful trajectories. The environments are designed for fast resets, easy replication, and scalable deployment, which makes them useful for RL training and evaluation. Strong browser-use agents still land at 69.3% average success, while vision-based approaches in the cited setup score 45.9% and 49.1%.

Why use WebArena?

• It pairs realistic web workflows with programmatic verifiers, so teams can measure task completion instead of relying on subjective review.
• The environments are self-hostable, which helps teams run experiments on their own infrastructure.
• Fast resets and easy replication make repeated agent runs practical for training and regression testing.
• The Infinity pipeline can generate environments in under ten hours at under $100 each, which lowers the cost of expanding a benchmark suite.
• Its tasks remain challenging for strong browser-use agents, so it can surface meaningful gaps in agent capability.

Who is WebArena for?

• AI researchers who need benchmark environments for autonomous web agents.
• ML engineers who want verifiable tasks for browser-use evaluation.
• RL teams who need scalable environments with fast resets and replication.
• Product teams studying long-horizon agent behavior across realistic web workflows.
• Benchmark authors who need programmatic verification for task correctness.

What are WebArena's key features?

What does WebArena integrate with?

• GitLab
• Wikipedia
• Google Maps
• Handshake
• Xero
• PayPal
• Elation EHR
• Gmail
• Superhuman

What are WebArena's use cases?

How does WebArena work?

1. Choose a benchmark suite such as WebArena, WebArena-Infinity, or VisualWebArena, then load the standalone, self-hostable web environment so your agent can start interacting with realistic sites.
2. Add the first task set and inspect the annotated programs to understand the expected end state, task structure, and what counts as a correct completion.
3. Run your agent against the environment and use Self Auditing plus Standalone Verifiers to check whether each trajectory satisfies the task requirements.
4. Reset the environment with Fast resets, then rerun experiments to compare models, prompts, or policies under the same conditions.
5. Scale the setup across more tasks and environments with Easy replication and Scalable deployment, so benchmark results stay consistent as your evaluation program grows.

Frequently asked questions