Atrex Kernel Agent workflow
An Alibaba framework for tuning GPU kernels with coding agents: a Supervisor checks correctness and speed first
- 212 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The agent edits and runs kernel code on GPU workers, spending compute time and tokens
- A campaign writes working files and profiling data into the chosen folder
Install
- Works in
Manual install
Clone `alibaba/atrex-kernel-agent`, open it in a coding agent and read `docs/quickstart.md`. Install the matching AKA packages, check the environment and launch a campaign on the chosen operator with the set number of iterations. Report the verified result and the output folder.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent edits and runs kernel code on GPU workers, spending compute time and tokens
- A campaign writes working files and profiling data into the chosen folder
This is third-party code. Review the repository files before installing.
What it does
Atrex Kernel Agent (AKA) drives GPU kernel optimization through a loop of research, editing, profiling and evaluation. A coding agent proposes candidates, while the Supervisor itself verifies correctness and compares the candidate with the best version on the same GPU allocation using an ABBA scheme. Only a verified improvement is accepted. The loop has five steps: measure the baseline, research and edit a managed draft with private evaluation tools, submit a report bound to a source snapshot and measurements, verify and promote, and continue with retained knowledge. Claude, Qoder, Codex, Pi and DSH agents are supported, with CUDA, Triton, CuTeDSL, FlyDSL and TileLang tooling for NVIDIA, AMD and T-Head PPU cards. Optimization knowledge lives in the GPU Wiki.
Who it is for. Performance engineers and GPU operator developers who want to hand optimization search to an agent but accept only measured improvements.
Good fit when
- An operator already works but slower than needed and a series of measured attempts must run
- An agent edit might break the result and is accepted only after a correctness check
- Work runs on several GPUs from different vendors and evaluation must happen on identical hardware
- Improvements must be kept with evidence so a next episode can build on them
Not a fit when
- There is no working Agate GPU environment: the main evaluation path runs through it
- A quick one-off code edit is needed: the loop with measurements is built for a series of iterations
- Something other than GPU code is being optimized, such as a web service: the tool is about kernels
Example request
Read docs/quickstart.md and set up the matching AKA packages from this checkout. Optimize /path/to/atrex-bench/operator on a <gpu_arch> Agate worker using CUDA, Claude as the coding agent, production mode, and up to 20 iterations.Limitations
An Agate GPU environment, prepared operator inputs and one of the supported agents are required. The version in the repository description is v0.4.0, an early project with a fast-changing architecture. The repository is large, about 1.6 thousand files and submodules, so clone it ahead of time. Part of the documentation, for example `docs/legacy-workflow.md`, describes the old process. The README does not estimate agent token cost for long campaigns. The latest commit in the repository is from November 2026.
How to disable. Delete the campaign working folder (`~/aka-opt` in the README example) and the repository clone. AKA packages installed into the environment are removed with the package manager you used.
FAQ
How do I launch a campaign?
Open the repository in a coding agent and give it the task: read `docs/quickstart.md`, install the packages, and name the operator path, GPU architecture, framework and iteration budget.
Which agents are supported?
Claude, Qoder, Codex, Pi and DSH. DSH has a separate adapter in `docs/dsh-harness-adapter.md`.
How is an improvement verified?
The Supervisor checks correctness and policy, then compares the candidate and the incumbent on the same GPU using ABBA. Only a verified improvement moves on.
Which cards and tools are supported?
NVIDIA, AMD and T-Head PPU, with CUDA, Triton, CuTeDSL, FlyDSL and TileLang tooling. Atrex-Bench evaluation runs through Agate.
What is the GPU Wiki?
A structured knowledge base on kernel optimization in the `gpu-wiki/` folder of the repository, serving the agent as a source of techniques.
Discussion
No questions or reviews yet.
Related
WorkflowsA Claude Code skill for GPU kernel tuning: the agent profiles, edits and benchmarks a kernel in a loop until speedup stalls.
ForgeKernel is a CLI that tunes a GPU kernel for hours: Claude Code writes variants, timing and profiling drop slower ones.
Holds write-ups of GPU kernel optimizations for Blackwell and Hopper so an agent answers from sources, not model memory.
A plugin for Claude Code and Codex with xdo, xdel and xflow modes: a main agent plans, a coder builds, an evaluator checks.
agentplay turns local Claude Code, Codex and Cursor sessions into a file map and a replayable step timeline.
Skills for writing, debugging, measuring and optimizing TileLang and CUDA programs from source, lowering and profiler evidence
README in short
The overview describes the five-step loop and the architecture diagram: baseline, agent, private tool runtime, GPU execution, acceptance. Release news, a quick start through an agent prompt and links to the documentation follow.