Browser Use has open sourced rl-env-use for people who know a domain skill a model should learn but not how to assemble the cases or the environment. A one line description of the skill is the input. Browser Use agents then mine real materials from the web, such as a registered trial protocol or a 10-Q, and the outcome published later. The tool writes tasks with gold answers computed in code, rewrites anything a frontier panel already solves, and rejects leaked, memorized, or ambiguous items. The set can be exported for an RL trainer or used to hillclimb a system prompt on a train and test split. The shipped example, trial-forecast, holds 35 forecasts from 29 ClinicalTrials.gov trials whose results were posted since July 2026, after the models' training data. Building it cost $72, including $17 on Browser Use. A hillclimbed Qwen 3.6 35B-A3B went from 0.53 to 0.78 on tasks it had not trained on, which the repo calls level with Opus 5.5. On the launch thread, Browser Use put that result on 16 held-out tasks at a climb cost of $8. Always answering 50 percent scores 0.75 on the same set, so most of the measured gain is calibration. The model stopped answering 0 or 1. The method follows Anthropic's workflow for automated eval design and hillclimbing, and the cases come from the live web. Claude Code, Codex, and similar coding agents can install the MIT licensed repo once a Browser Use API key and an OpenRouter key are set. Browser Use was founded by Magnus Müller and Gregor Žunič, who met as data science students at ETH Zurich. The company is based in San Francisco, and the mining agents run on its cloud browsers.