Launch: browseruse-agent-bench and LexBench-Browser

[Julia-Lex]

We are launching browseruse-agent-bench, a reproducible evaluation framework for browser agents, together with LexBench-Browser, the built-in public dataset used by the default benchmark workflow.

LexBench-Browser v1.0 includes 210 no-login tasks across 107 real websites, with Chinese and English instructions, reference steps, key points, common mistakes, scoring rubrics, and robustness tags.

What you can do:

• Run browser-use, Skyvern, Agent-TARS, and other browser-agent integrations.
• Compare agent, model, browser backend, and evaluation strategy under a shared workflow.
• Inspect trajectories, cost, latency, token usage, and task-level evaluation details.
• Submit reproducible results, new agents, and dataset task proposals.

Useful links:

• Repository: https://github.com/lexmount/browseruse-agent-bench
• Dataset: https://huggingface.co/datasets/Lexmount/LexBench-Browser
• Documentation: https://docs.bubench.lexmount.io/
• Evaluation protocol: https://github.com/lexmount/browseruse-agent-bench/blob/main/EVALUATION_PROTOCOL.md
• Result submissions: https://docs.bubench.lexmount.io/en/leaderboard/result-submissions

For official result coordination, dataset questions, and collaboration requests, contact lexbench@lexmount.com.

We welcome feedback from browser-agent builders, benchmark users, and teams who want their agents or results reviewed for the public leaderboard.