Autoresearch

When a human writes high-level instructions in program.md, the agent repeatedly loops through modifying train.py, running a five-minute training job, and checking the results. It uses Git as research memory: if the metrics improve, it commits the change; otherwise, it rolls back with git reset. It even independently discovered that improving iteration speed was more beneficial than increasing batch size. Beyond ML training, this approach can be applied to any coding task that can be measured with metrics.