WebArena is a benchmark for developing and evaluating AI agents that use websites—not a search engine or directory for discovering websites. It gives agents realistic, multi-step tasks in functional website replicas, then checks whether they found information or completed the requested actions.
What is WebArena?
WebArena is a self-hostable environment for testing language-guided agents on web tasks. Its ICLR 2024 paper describes functional website replicas across e-commerce, social forum discussions, collaborative software development and content management, alongside tools such as a map and external knowledge resources such as user manuals. The project is intended to make agent performance measurable under shared conditions.
As an Amazon Associate I earn from qualifying purchases.
The benchmark focuses on translating a high-level instruction into concrete browser interactions. For example, an agent may need to navigate a site, locate relevant information, or change something on the site rather than simply produce a plausible-sounding answer.
How does WebArena test web-browsing agents?
WebArena evaluates two broad kinds of work. For information-seeking tasks, the project describes comparing an agent’s predicted answer with a reference. For action tasks, it can check whether intermediate site states have expected properties. This lets an evaluation distinguish an answer that merely sounds right from an interaction sequence that reaches the intended outcome.
#1 Best Overall
- Information retrieval: The agent answers a question using information available in the environment, and its answer is compared with a reference.
- Action execution: The agent performs interactions, and the resulting site state can be checked against expected conditions.
Because tasks can involve multiple steps, success depends on more than fluent language. Navigation, interpreting the instruction, interacting with the right controls and reaching the required state all matter.
What do WebArena’s published scores show?
The ICLR 2024 paper reported a 14.41% end-to-end task success rate for its best GPT-4-based agent and 78.24% for human performance in that paper’s experimental setup. The authors’ abstract described complex tasks as challenging and reported the gap between the agent and human results.
A later, separate comparison appears on OpenAI’s 2025 Computer-Using Agent page: its browser-use table reports 58.1% for OpenAI CUA, 36.2% for the previous state of the art and 78.2% for human performance on WebArena. These are figures from OpenAI’s comparison, not a continuation of the ICLR paper’s experiment. The available sources do not establish that the evaluations used identical protocols, task sets, interfaces or scoring details, so the figures should not be read as a directly controlled before-and-after comparison.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →| Evaluation | Agent result | Comparison result |
|---|---|---|
| ICLR 2024 paper | 14.41% for the paper’s best GPT-4-based agent | 78.24% human end-to-end task success |
| OpenAI 2025 browser-use comparison | 58.1% for OpenAI CUA | 36.2% previous state of the art; 78.2% human performance |
Scores are useful only with their evaluation context attached. When comparing results, identify the evaluation and date, the agent and browser interface, what counted as success, and which human or prior-system comparison came from the same report.
Rank #3
Can you run WebArena yourself?
The canonical WebArena repository identifies itself as the implementation for reproducing the paper’s results, and the official project site links to code, data, a Docker environment and leaderboard resources. Those links establish that the project provides reproducibility resources, but they do not by themselves confirm that every linked service is currently live.
A notice in the repository dated December 5, 2024 recommends AgentLab for updated experimentation infrastructure. It describes parallel experiments using BrowserGym, integration of benchmarks including VisualWebArena, unified leaderboard reporting and improved handling of environment edge cases. Treat that as dated guidance: check the repository’s current instructions before choosing a setup path.
Rank #4
What WebArena makes competitive—and what it does not
WebArena makes agent performance comparable by giving systems shared tasks in a fixed web environment and evaluating task outcomes. “Competitive” here refers to measurement across agents, not competition among websites. It is not a consumer website-discovery service, a search engine or a website-ranking product.
The ICLR 2024 paper lists Shuyan Zhou and eleven coauthors. The reviewed project materials do not establish that the person behind a first-person headline personally built WebArena; the benchmark should therefore be described as the research team’s work unless author-specific evidence supports an individual claim.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
