24 hours, 1,000 tool calls: what the kernel case study proves
Benchmarks measure an agent’s sprint. Meta’s kernel-optimization study measures its marathon — and it’s the more interesting number.
The setup
Meta tested Muse Spark 1.2’s ability to iteratively optimize GPU kernels over more than 1,000 tool calls, in runs lasting up to 24 hours. Inside Muse Code’s agentic environment, the model writes a kernel, compiles it, profiles it against a provided baseline, and uses the profiling data to try again — a tight write-compile-profile-improve loop sustained for a full day. The benchmark targets were KDA and MLA attention kernels on NVIDIA Hopper GPUs.
One rule made the test honest: models were prohibited from importing third-party kernel libraries like FLA. No wrapping an existing fast implementation and taking credit — the agent had to apply genuine kernel-optimization knowledge and implement the algorithms in Triton itself.
What the agent actually built
The solutions weren’t generic. For KDA — benchmarked against the FLA Triton implementation — Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, combining standard fusion and tiling with KDA-specific moves like re-centering the gated cumulative decay at the chunk midpoint. For MLA — against a PyTorch reference at batch size 1, 64 heads, sequence length 8192, latent dimension 512 — it designed a two-kernel Triton pipeline that reuses the shared KV latent as both K and V.
Those are the kinds of optimizations a specialist writes after understanding the algorithm’s structure, not surface-level loop tweaks. And the agent kept finding substantial improvements over the baseline as the run went on — progress that accumulated across hundreds of iterations rather than plateauing after the first burst.
Why 24 hours is the real headline
Plenty of models can improve a kernel for twenty minutes. Sustaining directed progress for a day requires infrastructure most agents don’t have. Three pieces carry the load. Goal conditioning keeps iteration number 700 aimed at the same objective as iteration one — the model was trained for exactly this. Context compaction retains the knowledge that matters (which strategies failed, what the profiler said) without drowning in a day’s worth of history. And the event log makes the runtime restart-safe — over 24 hours, something will hiccup, and a crash that erased hour 23 would make the whole exercise pointless.
What it signals for ordinary work
Few teams optimize Hopper kernels. But the shape of the task — iterate against a measurable target, learn from each attempt, don’t lose the thread — matches a lot of unglamorous engineering: chasing a flaky test suite, grinding a bundle size down, migrating an API a hundred call sites at a time. The case study is Meta’s evidence that “kick it off and check tomorrow” is a real workflow now. The evaluation methodology is published on Meta’s research site; the product side of long-horizon work starts in the docs.