Autoresearch as Universal Optimization Primitive: Pattern Expands from ML to Any Single-Metric Domain

paddo.dev·medium signal

A March 15 paddo.dev analysis documents how the autoresearch loop (agent modifies code → evaluate single metric → git tracks → loop) has become a general-purpose primitive beyond ML training. Documented deployments include GPU kernel optimization (~40 experiments/hour via Autokernel), Apple Silicon MLX optimization (val_bpb 2.667→1.294), test coverage automation, bundle size reduction, and agent-on-agent systems where one agent optimizes another's code. The key design insight: the constraints (single file, one metric, fixed budget, git memory) are what make the loop portable and reliable.