Autoresearch, Claude and Constrained Optimization

The author conducted an experiment using Claude AI to develop a file compression algorithm without human supervision, setting specific constraints (perfect decompression and maximum 300-second runtime) to test how well AI agents can tackle real-world optimization problems. The project was designed to measure whether AI can reliably solve constrained problems that require testing, iteration, and decision-making—rather than simple gradient-based optimization—and to benchmark the results against existing compression tools. This proof-of-concept explores whether AI-generated solutions could eventually replace external dependencies and libraries for specific tasks.