MaxKernel
A multi-agent system that generates and tunes high-performance TPU kernels, evaluated with Gemini 3.1 Pro and released under Google's accelerator-agents repository.
The paper describes HITL, fully autonomous, and graph-based search paradigms sharing planning, implementation, debugging, testing, and profiling sub-agents. Experiments on JaxBench and real model workloads were run with Gemini 3.1 Pro. Code is published in the maxkernel directory of AI-Hypercomputer/accelerator-agents.
Human involvement
Person directed, agent wrote the code. The system includes a human-in-the-loop agent as well as autonomous loops. The paper is a research artifact; the degree of hand-written kernel code in the repo is not quantified here.
Evidence
- Paper
All large language model queries and agent interactions throughout our experiments were conducted using the Gemini 3.1 Pro model. Code: https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel
- Official repository
Notes
The paper URL uses MaxKernel (capital M); the GitHub tree that returned 200 is maxkernel in lowercase. The directory listing included adk and claude-code helper folders; the evaluated model in the paper is Gemini 3.1 Pro.