- The author conducted an experiment using an AI agent (Claude Code) to autonomously develop a file compression algorithm, aiming to test its ability to optimize software with measurable objectives and constraints.
- The problem was framed with clear goals: minimize compressed file size while ensuring lossless decompression and a time limit of 300 seconds for compression/decompression, using Rust for implementation and a dataset of varied file types.
- Over ten iterations, the AI improved compression by implementing techniques like LZSS and entropy checks, achieving notable results, especially for audio and video files, though it consistently optimized only for compression without considering other factors like speed.
- Key learnings include the AI's tendency to quickly consider tasks 'done', highlighting the need for explicit looping mechanisms; the critical role of objective function design in avoiding over-optimization; and challenges in applying this to real-world metrics with slower feedback loops.
- Limitations noted include model variability, cost (about $4 per iteration), hardware dependencies, and the difficulty of choosing optimization functions, underscoring that while promising, autonomous agent research requires careful setup and metric selection for practical use.