Argon Speed
Argon Speed is a GPU kernel engineering team made of five Gemini 4 Argon agents.
Core trick
Agents write CUDA or Triton kernels, then an independent judge times each one on an H200 against SGLang's production kernel.
Tags
Description
Argon Speed is a GPU kernel engineering team made of five Gemini 4 Argon agents. It runs around
the clock on Gemini Managed Agents, with no laptop involved and no GPU of our own. Each agent owns
one problem in the Open Kernel Challenge: the kernels that serve Muse Spark Open, an
816B-parameter mixture-of-experts model, on SGLang.
Each run, an agent:
- reads SGLang's source at the judge's exact commit, the judge's per-shape profiles and its own lab
notebook;
- designs a change and writes it in CUDA or Triton;
- compiles it in its sandbox;
- sends it to the challenge's independent judge, which runs hidden tests and times it on a real
H200 against SGLang's own production kernel.
An hourly trigger wakes each agent, and every run leaves notes for the next one. Guardrails keep
it honest: a check budget, the judge's token never enters a sandbox, and only a person can make a
scored run.
In its first afternoon it beat SGLang on all five problems and got onto all five public
leaderboards. Scored runs are 1.02x to 1.32x faster than SGLang. It read 58M tokens to get there,
for $18.
The point is where this goes next. Aimed at a model's own serving stack, every speedup makes the
next iteration cheaper and faster, so the rate of improvement compounds. That's the speed in Argon
Speed.
README excerpt
# Argon Speed **AI that makes itself faster.** Five Gemini 4 Argon agents work around the clock on Google's Managed Agents. Each one optimizes one GPU kernel of a frontier model's serving stack. Today that stack is Muse Spark Open's, an 816B-parameter MoE served by SGLang, and the kernels are the five problems of the [Open Kernel Challenge](https://openkernelchallenge.dev). The challenge's independent judge measures every result on a real H200 against SGLang's own production kernels. Nobody's laptop is involved, and we have no GPU of our own. Point the same loop at Argon's own serving and every speedup makes the next iteration cheaper and faster, so the rate of improvement compounds. That's the speed in Argon Speed. **Pitch:** [slides (PDF)](pitch/argon-speed.pdf) · [2-minute narrated video on YouTube](https://www.youtube.com/watch?v=u9LV2i_eBSw) [](https://www.youtube.com/watch?v=u9LV2i_eBSw) ## How we use Gemini 4 Argon: million-token engineering - **The whole stack in context.** Each run gives Argon a 1.5M-token budget for: - SGLang's source at the judge's exact commit; - the challenge's spec, tests and judge environment; - the judge's per-shape profiles with speed-of-light bounds; - its own lab notebook from earlier runs. So far that's 58M tokens over 32 hosted runs, 94% of them served from cache. - **Engineering, not chat.** Argon runs the whole loop itself: - designs a change and writes the CUDA