Every team that ports GPU code underestimates it, and the underestimate always lands in the same place. The mechanical translation is largely solved. The expense is the long tail: kernels that compile and run and produce almost the right answer, kernels that are correct but several times slower, and kernels that fail only on one input shape. That tail is where months go.

This talk presents a method for putting an LLM agent to work on that tail without trusting it. The centerpiece is the verification harness, not the model. Numerical equivalence gates with tolerances chosen per operation. Shape and dtype sweeps that catch the cases a human reviewer skips. Performance regression gates so a successful port cannot quietly be a slowdown. We walk through a real migration of vision and control kernels to HIP on an embedded GPU, categorize every failure the agent produced, and are specific about which categories it handled reliably, which it handled with review, and which it got wrong in ways that looked plausible. Attendees get a harness design they can rebuild