A developer tested whether Claude Code with Sonnet 4.6 could autonomously improve a file compression algorithm through iterative, unsupervised optimization. The experiment, hosted at github.com/smitec/agent-compression, framed compression as a constrained optimization problem: minimize final file size while maintaining bit-perfect decompression and a 300-second timeout per file. The scaffold was built in Rust to enforce function signatures via the type system, starting with a stub that copied bytes without compression. A benchmarking suite tested against public domain video, audio, text, and random data files up to 150MB, measuring combined compression ratio as total compressed bytes divided by total original bytes.

Over 10 iterations spanning two weeks, the agent produced a custom LZSS implementation in the first pass, then spent nine iterations adding entropy checks and encoding techniques. Each iteration averaged $4 USD in API costs under default settings. The model consistently formed a single hypothesis per iteration, implemented changes, ran benchmarks, and declared completion without further prompting. The author cleared context before each run and accepted plans without modification, aiming for fully autonomous decision-making.

A final extended benchmark compared the evolved algorithm against standard compression tools on a fresh dataset to control for data-specific overfitting. The exercise was not designed to beat production compressors but to evaluate whether AI agents can reliably tackle quantified engineering problems with hard constraints. The setup mirrors real-world scenarios where teams optimize metrics like latency or cost under strict correctness requirements.

The experiment demonstrates that current models can execute multi-step optimization loops with measurable outcomes, though the single-hypothesis-per-iteration pattern suggests limitations in exploratory search. For engineering teams, this approach offers a template for automating well-defined optimization tasks — such as database query tuning or serialization format selection — where objective functions and constraints are explicit. The cost per iteration and time investment remain practical for targeted problems, but the lack of multi-path exploration may require human intervention for complex search spaces.