A new tool called Researcher Skill transforms AI coding agents into autonomous scientists capable of running dozens of experiments overnight. The open-source project by krzysztofdudek packages everything into a single markdown file that agents like Claude Code and Cursor can read and execute.

Once dropped into a codebase, the skill interviews developers about optimization goals, then sets up a dedicated git branch to run experiments autonomously. It systematically tests hypotheses, commits successful changes, and reverts failures while maintaining detailed logs. The approach mimics scientific methodology with iterative testing, measurement, and analysis.

What makes this compelling is its breadth of application beyond machine learning. The tool works on API latency optimization, test suite performance, bundle size reduction, prompt engineering, algorithm tuning, and configuration optimization. Each experiment follows a consistent pattern: formulate hypothesis, implement change, measure results, and decide whether to keep or discard. The system maintains experiment history in an untracked .lab/ directory that survives all git operations, creating persistent knowledge that outlasts individual coding sessions.