Agent teams that restructure themselves beat coding agents at GPU tuning
A new paper has a supervisor rebuild a team of coding agents when they stall, and reports faster GPU code than expert-written kernels on most tested workloads.
The short version
- Existing coding agents stall at local optima on expert-written GPU kernels such as FlashInfer, the authors found.
- A central supervisor diagnoses each stall and restructures the team of agents.
- A persistent file system holds a learning document, an evolution history and per-team artifacts.
- With an extended token budget it wins on 36 of 47 FlashInfer-Bench production workloads; the other 11 land at 0.86-0.99x.
- Code is public on GitHub; the numbers are the authors' own and not independently checked.
Why it matters
Agents are usually shown on routine coding. This paper tests them on work where expert humans have already tuned the code, which is a harder bar. If the results hold up, how a team of agents is organized may matter as much as how strong the underlying model is.
Where coding agents get stuck
A GPU kernel is a small program that runs on a graphics chip, and AI systems spend much of their time executing them. Making one faster is open-ended work, with no single right answer to aim for. As arxiv.org reported in this paper's abstract, the authors found that existing coding agents tend to settle into local optima on real, expert-written kernels. A local optimum is a point where small tweaks stop helping even though a much better design exists.
A supervisor that rebuilds the team
The authors' answer is called Evolving Agent Teams. A central supervisor looks at why the current team has stalled and then restructures it, rather than letting the same agents keep trying the same moves.
The supervisor is also helped by memory. A persistent file system holds a learning document, an evolution history and artifacts for each team, so each restructuring draws on what earlier teams tried.
What the authors report
With a matched token budget (the same amount of model text processed), the system beat strong coding-agent baselines. With a larger budget, it wrote a CUDA kernel from scratch that reached a 1.2x mean speedup over FlashInfer's production MLA paged decode kernel. It won on 36 of 47 production workloads and scored 0.86-0.99x on the other 11.
The abstract lists other results too:
- 3.01x over vLLM on KDA prefill on AMD MI350X
- 3.44x over FlashKDA on NVIDIA B200
- 30.3 tokens/s decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro
- 62.1 on RoboDojo robot manipulation with held-out seeds, against 55.3 for the best leaderboard entry
How to read the claims
These are the authors' own figures, taken from an abstract. We have not checked the methods or tried to reproduce them. The extended-budget results also use more compute than the matched-budget comparison, so cost matters when judging them.
The code is public on GitHub, so others can test the claims.
What to watch
- Will independent groups reproduce the speedups on kernels the authors did not pick?
- How much extra compute does the extended-budget run cost compared with the gains?
Try it
Open the abstract and look at the matched-token-budget comparison first, then check the GitHub repository to see which models and hardware you would need to reproduce it.
Sources and links
- Reportingarxiv.org: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by arxiv.org · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.