Agentic Kernel Optimization
If you find our work useful, please consider giving us a star 🌟
Achieve competitive GPU kernel performance in just hours.
- AKO4X beats the FlashInfer expert on 9 of the 10 operator families profiled on NVIDIA B200 — up to 30.71× on DSA sparse attention, from the MLSys-2026 FlashInfer-Bench contest set.
- AKO4ALL, the drop-in skill, beats the expert on all 4 inference operators — one prompt, ~1h per kernel.
Full results on the project page →
- 📄 [2026.05.31] The AKO tech report is now available.
- 🚀 [2026.05.31] AKO4X is now open-source — the closed-loop, campaign-based system behind our MLSys 2026 competition entry.
- ✨ [2026.05.31] AKO4ALL is now a single drop-in Claude Code skill — invoke it in any working directory.
- 🚀 [2026.03.24] AKO4ALL is released. Check out the project page.
AKO is not a new agent or model — it is a harness (optimization environment) for existing coding agents such as Claude Code. It places the agent into a well-structured environment where the evaluation criteria, benchmarking tools, profiling interfaces, and optimization trajectory are all clearly defined and readily accessible.
| Tool | Description | Repo |
|---|---|---|
| AKO4ALL | Single drop-in Claude Code skill — open and minimal, for any kernel and any language. Bring your own benchmark or use the built-in KernelBench evaluator. | TongmingLAIC/AKO4ALL |
| AKO4X | Advanced, eXtensible harness: single manual sessions or closed-loop multi-round campaigns with cross-run memory, master/sub agent separation, and opt-in harness co-evolution. Benchmark-swappable via a thin adapter (default flashinfer-bench). | TongmingLAIC/AKO4X |
The AKO tech report is now available.
We would like to thank the following open-source projects that inspired and supported the development of AKO:
- KernelBench — for the benchmark and evaluation format used by AKO4ALL's built-in evaluator.
- FlashInfer — for the LLM inference kernel library and the flashinfer-bench benchmark infrastructure on which AKO4X is built.
- autoresearch and autokernel — AKO's design was inspired by their work on autonomous optimization loops.
We also thank Modal for the GPU credits that powered our MLSys 2026 competition runs.
If you find AKO useful, please cite:
@misc{ako2026,
title = {{AKO}: Agentic Kernel Optimization},
author = {Shuxiao Xie and Shuyang Xie and Dezhi Ran and Wei Yang and Tao Xie},
year = {2026},
howpublished = {\url{https://tongminglaic.github.io/AKO}},
note = {Technical report}
}