Intel / article

CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language

Source: arxiv.org
Author: Qi Fan, An Zou, Yehan Ma
Date: Published Sep 2, 2026

Context

CUDA-Harness separates understanding a requested GPU calculation from writing and testing its implementation. It first produces a structured specification of inputs, outputs and operations, then generates the CUDA kernel, the code that performs the parallel computation.

Its test generator works in a separate from the code generator, which is intended to stop tests from being tailored to whatever the implementation happens to do. Validation then guides repairs before speed optimization. The separation has a limit the authors document: when the request itself is misread, the misreading can carry into both the kernel and its synthesized tests. Prompt misinterpretation accounts for 38.1% of the cases where the synthesized tests passed but the ’s own check failed.

Terms in this piece · Glossary

agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source arxiv.org