An orchestrated agent system for production-grade GPU kernel development

---

Overview

Atrex Kernel Agent (AKA) turns an evaluator-owned operator into a measured, optimized GPU kernel. It coordinates coding agents, GPU profiling, correctness checks, performance verification, Git isolation, recovery, and final packaging while keeping acceptance and termination under mechanical supervisor control.

The optimization application is launched through the installed aka optimize command. New launches default to the Light profile: native Agent execution with session capture, managed candidate workspaces, a private HTTP Tool Runtime and Supervisor-owned acceptance. Use --profile light-sandbox for Bubblewrap, or --profile workflow for the Legacy workflow. Core selects the implementation, and the installed Workflow module supplies its episode engine. The supervisor establishes an evaluator-backed V0 and, when enabled, a framework-native V1. Coding agents then explore unified Light Episodes (Legacy retains fast/full/goal modes). The supervisor owns correctness, production policy, performance verification, and promotion; every outcome is recorded without letting an unverified candidate replace the incumbent.

AKA supports:

• SOL-ExecBench and native Atrex-Bench operator layouts;
• NVIDIA, AMD, and T-Head PPU (zwm890p) targets through isolated sandbox execution;
• gateway and Bubblewrap-isolated OpenSSH GPU execution, with automatic environment recovery;
• Triton, CuteDSL, CUDA, FlyDSL, and TileLang campaigns;
• Claude, Qoder, Codex, and Pi coding-agent backends;
• leaderboard and fail-closed production modes;
• automatic local plugins and structured GPU Wiki retrieval with traceable attribution;
• production-mode numerical probes and mode-specific performance verification;
• resumable, Git-isolated optimization with canonical measurement history.

For Agate evaluations requesting two or more GPUs, production source-policy inspection is disabled automatically. The GPU count comes from the evaluator's benchmark_contract.distributed_evaluation (torchrun, nccl, world_size >= 2), the same contract used to submit Agate's num_gpus. These campaigns can reuse communication libraries and the validated V0 without a self-authored framework conversion. Correctness, numerical probes, performance checks and immutable evaluation inputs remain enforced. Single-GPU and SSH campaigns retain source-policy inspection. Invalid distributed contracts are errors; operator names and candidate dependency declarations do not enable this exception. The source-policy decision is frozen per campaign from private evaluator metadata, or from the committed V0 metadata on resume (the original operator directory before V0 exists). Creating or editing metadata in the candidate workspace cannot change that decision. Legacy metadata without a distributed_evaluation key retains single-GPU submission and source checks even if hardware.required_gpus describes multiple GPUs. An explicit distributed contract requires every supplied hardware.required_gpus value, including 1, to match its world_size.

ATREX_SANDBOX_ATREX_BENCH_VERSION and ATREX_SANDBOX_REQUIREMENTS_FILE can pin an Agate runtime. The requirements path is resolved through the Problem plugin, validated as UTF-8 flat package specifications, and frozen into the Task snapshot. Baseline, Agent sandbox, dispatch and resume bind the same frozen file; an explicit empty file remains empty. When replaying a typed request through the CLI, its parsed contents must match the request, so missing or changed pins fail instead of silently selecting a different runtime.

News

• [2026-09] From leaderboards to real-world workloads: accelerating Qwen3.8-Max's core inference operators (WeChat): AKA delivers lower average latency than expert implementations for all three key operator families across 60 real-world shapes, with up to 1.55× speedup over production/official baselines.
• [2026-09] We released Atrex Kernel Agent v0.3.0 with resumable fast, full, and goal episodes; native Atrex-Bench and multi-framework production campaigns; Claude, Qoder, Codex, and Pi backends; isolated SSH GPU execution with recovery; PPU profiling; plugin-backed GPU Wiki retrieval; and production-mode supplemental numerical probes. See the v0.3.0 release notes.
• [2026-08] We slimmed down Atrex Kernel Agent by consolidating on a single orchestrated workflow and removing legacy paths and redundant context for a smaller context footprint and lower token usage.
• [2026-07] We helped Qwen3.8 rank No. 1 on the SOL-ExecBench FlashInfer operator optimization leaderboard.
• [2026-07] We released Atrex Kernel Agent v0.2.0 with an orchestrated clean-session loop, native SOL-ExecBench operator workflow, Triton-to-Gluon conversion support, and a fuller NVIDIA profiling toolchain. [Release]
• [2026-07] We released the Atrex paper: Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent.
• [2026-06] We released Atrex Kernel Agent v0.1.0 as the initial open-source version, with the GPU Wiki knowledge base, profile-driven optimization workflow, profiling tools, and reference templates. [Release]

Quick Start

See the Quick Start guide for prerequisites and complete runnable examples of the orchestrated optimization loop.

Or start a coding agent such as Claude Code, Codex, or Qoder in this repository and ask it to launch an AKA optimization task. We recommend the following prompt:

Use the installed `aka optimize` command with this checkout as `--repo-root` to start one optimization task for atrex-bench/xx. Put the workspace under ~/aka-opt, set the platform to H20, use the local sandbox, use claude as the Agent CLI, set max-iters to 300, specify cuda as the framework, and run in production mode.

Documentation

Document	Contents
Quick Start	Setup, commands, campaign steps, configuration, and outputs
Architecture Design	Components, authority boundaries, state machine, verification, and recovery
GPU Wiki	Structured hardware/kernel knowledge, queries, and trace mining
Local plugins	Extend AKA tools and Skills through automatically discovered local plugins
Light modules and launch profiles	Default Light workflow, explicit Legacy selection, sandboxing and optional module configuration
Optimization dashboard	Optional Core plugin for local progress, performance history and settled token usage
v0.3.0 release notes	New features, correctness and measurement changes, and upgrade guidance

Run aka optimize --help and aka optimize --repo-root "$PWD" -- --help for the framework and optimization interfaces and defaults.

For a GPU server reachable through OpenSSH, use a dedicated low-privilege account and pass --sandbox-ssh user@gpu-host --sandbox-ssh-gpu 0 with an explicit --framework. AKA keeps Agent, Git, memory, and episode state local, transfers only the sandbox allowlist to a fresh remote temporary directory, and runs it in a mandatory networkless Bubblewrap namespace before copying back requested artifacts. Use --sandbox-ssh-runtime-bind for read-only venv or toolchain trees outside the minimal system mounts. Runtime sources are checked against sensitive host paths after remote symlink resolution, and only the assigned physical NVIDIA GPU is visible; MIG execution fails closed until capability-node assignment is supported. If the remote environment fails its GPU health probe, AKA stops, preserves the active worktree, and starts a detached monitor that resumes the original command after the server recovers. See Quick Start: Isolated OpenSSH GPU host. The generated stop-recovery.sh also stops a recovered optimizer and leaves automatic recovery disabled until the generated recover.sh is run explicitly. The recovery monitor stays with the restarted optimizer: an unexpected exit returns to health polling and automatic restart, while only a durably clean zero exit completes recovery.

Acknowledgements

AKA builds on and learns from many open-source projects, including:

• GPU kernel projects: CUTLASS, cutex, cuLA, FlashAttention, FlashInfer, FlyDSL, Triton, DeepGEMM, LeetCUDA, FlashMLA, Composable Kernel, cute-gemm, hpc-ops, AIter, quack, and TileLang.
• Knowledge and agent tooling: KernelWiki, modern-gpu-programming-for-mlsys, ncu-report-skill, humanize, AKO4ALL, and KDA.

Citation

If AKA is useful in your work, please cite the Atrex paper:

@misc{atrex2026,
  title         = {Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent},
  author        = {Lingyun Yang and Yuxiao Wang and Shenghao Liang and Linfeng Yang and Daocheng Ying and Chunbo You and Rui Zhang and Luping Wang and Yinghao Yu and Guodong Yang and Liping Zhang},
  year          = {2026},
  eprint        = {2607.14541},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2607.14541}
}

Application framework

After installing the matching framework, legacy and light packages, run:

aka optimize --repo-root /path/to/aka -- \
  --op-dir /path/to/problem --platform L20N --framework Triton

--repo-root selects the matching AKA checkout; --op-dir selects the native Atrex-Bench or SOL problem. Omitting --profile selects native Light, not Bubblewrap. Wiki, Direction/Experiment recording and generated memory remain disabled by default. For a new Legacy campaign, add --profile workflow before --; existing recorded continuations retain their original selection. Do not reuse a Legacy workspace as a Light workspace. See the installation and application selection guide and Core interface.

License

Licensed under the Apache License 2.0.