browser-use/rl-env-use is a tool that builds reinforcement learning (RL) environments and evaluation tasks in any domain, based on real cases a browser agent collects from the web. An RL environment is a framework in which an agent repeatedly acts and receives a reward signal based on the outcome; rather than designing that framework directly, this repository derives its tasks from real cases on the web. There are no public releases, and the repository was created on October 21, 2026.
The README describes a three-part procedure: it computes the ground-truth answer to a task in code, filters out tasks an agent could already solve or whose inputs are ambiguous, and provides a procedure for iteratively improving the prompts used for training. The repository includes a clinical trial outcome prediction case, and the README states that 35 prediction tasks were constructed from 29 trial documents. The license is MIT, the displayed primary language is HTML, and the star count is 52.
However, the only real case confirmed in the README is clinical trial outcome prediction, so it is hard to tell from the documentation alone whether the same procedure holds up in other domains. The performance figures in the README are evaluation results presented by the repository, not independently re-verified values. Since there are no public releases, it is also difficult to trace version-by-version improvement history. Judging it on performance is therefore premature, but it is a suitable repository for internalizing the idea of verifying a task's answer in code, or for use as an example case.