An honest little system, which is rare enough to note. LangGraph agent loops over CUDA kernels: propose a change, compile via NVRTC through the driver API, validate against a reference with NumPy, time 100 launches after 10 warmups with CUDA events, keep the fastest that passes every case. Ranking is the geometric mean of latency across performance cases; correctness cases gate but don't score. Optionally it reads Nsight Compute counters and NVIDIA docs between attempts. The limits section deserves credit because it's where most similar repos hide the bodies: no cuBLAS comparison, a generated reference is not an oracle, passing your cases proves nothing general, and generated input scripts run locally with no sandbox. That last one means the agent's Python output executes on your machine by design, so read your own `--input-cases` before pointing this at anything you care about. Developed on an RTX 3060 Laptop GPU with Python 3.12 and an OpenAI key, default model gpt-5-mini. The results are two GEMM runs on that same laptop GPU, which tells you the honest scope: this is a promising harness, not a cuBLAS replacement. The open question: does the agent actually discover anything a decent CUDA programmer wouldn't, or does it mostly walk the occupancy spreadsheet? Run 023 exists because run 022's answer wasn't good enough on its own inputs, which is suggestive either way.