AKO4X
An agent-driven framework that automates GPU kernel optimization by pointing a coding agent, currently Claude Code, at a kernel operator and letting it iterate inside an isolated workspace using profilers and per-language knowledge the framework provides.
AKO4X spawns a fresh, isolated workspace for a target kernel operator, such as an attention or GEMM shape, and lets a coding agent iterate on it using tools, profilers and per-language knowledge the framework supplies. It can be run as a single session, where one workspace is spawned and the agent optimizes until done, or as a multi-round campaign, where a persistent master agent autonomously spawns rounds back to back, archives each round's best variant, and seeds the next round from the growing archive.
Three parts of the system can be swapped or extended: the benchmark that scores a kernel's performance, which defaults to flashinfer-bench; the per-language knowledge and tool-wrapper "SKILLs" layer, added by dropping a folder into a templates directory; and, in an advanced operating mode, the harness itself, where the agent proposes improvements to its own templates each round and a master process gates them in. A companion plugin makes an optimization run watchable and steerable live for DeepSeek Harness.
It is written in Python and reports a competition entry result on contest kernels against a concurrent industry submission.