VibeKernel: Long-Horizon GPU Kernel Optimization with General-Purpose Coding Agents

GPU kernel optimization is highly iterative yet straightforward to evaluate, which has motivated systems that place language models inside specialized, kernel-specific search procedures. We ask whether a general-purpose coding agent can do the same work without such a loop. We present VibeKernel, an environment that gives an off-the-shelf coding agent a persistent kernel workspace and controlled tools for reading, editing, building, benchmarking, and profiling kernels, while leaving the optimization strategy to the agent. On a multi-head latent attention kernel written in Triton and evaluated over $47$ workloads, VibeKernel with Claude Code reaches a $299\times$ geometric-mean speedup over a naive PyTorch reference, ahead of the specialized workflows K-Search ($257\times$) and AdaExplore ($178\times$) under the same model and cost budget. On the public SOL-ExecBench leaderboard, a kernel produced this way ranked fourth at the time of writing. We then characterize two long-horizon failure modes: commitment to the incumbent implementation and the promotion of speculative conclusions to established facts during context summarization. We address both with a version-control-inspired experiment memory that preserves measured kernel snapshots across clean context resets while discarding the associated reasoning. Over a long optimization horizon, compaction-based continuation plateaus at $227\times$, whereas the version-control-inspired memory sustains improvement to $289\times$.