Optimizes code performance and software metrics through autonomous, iterative experimentation loops.
Autoresearch is a specialized capability for Claude Code that automates the process of performance tuning and code optimization. It functions as a continuous experiment loop that generates hypotheses, applies code modifications, runs benchmarks, and evaluates results against a primary metric. By operating within a secure git worktree, the skill safely iterates on your codebase—keeping improvements and discarding regressions—until your optimization goal is reached or all potential avenues are exhausted. It is ideal for developers looking to improve execution speed, reduce memory footprint, or minimize API costs without manual trial-and-error.
Key Features
01Adaptive search strategies including Explore, Exploit, and Ablation
020 GitHub stars
03Detailed experiment logging and session state persistence in JSONL
04Autonomous hypothesis generation and iterative testing loop
05Automated benchmarking with metric-driven keep/discard logic
06Strict git worktree isolation for safe, sandboxed experimentation
Use Cases
01Optimizing function execution time or reducing memory overhead
02Reducing LLM token usage or cloud infrastructure costs through code tuning
03Iteratively refactoring code to meet specific performance benchmarks
