The agent workflow behind #1–3 at the MLSys Kernel Contest

What it does — KDA is not a library or a harness — it’s a documented workflow for coding agents doing performance-sensitive CUDA work. The core artifacts are a generic starter prompt (prompts/basic-flow.md), repo-facing agent instructions (CLAUDE.md), and two pinned skills: KernelWiki and ncu-report-skill. An agent running this flow drafts a plan, implements in small iterations, verifies after each meaningful change, and logs candidates, benchmark results, profiling evidence, and promotion decisions as it goes. As shipped it’s Claude Code-shaped — a skills directory plus an optional humanize planning plugin — though the workflow is deliberately independent of any single benchmark harness or hardware target.

The interesting bit — The discipline is the product. Kernels are treated as candidates that must earn promotion through validation and profiling evidence, and the prescribed workspace (candidates.jsonl, benchmark.csv, profile/) exists so another engineer can reconstruct what was tried, what passed, and why the winner won. The credibility claim is concrete, too: HAN Lab Mafia’s solutions ranked #1–3 on tracks at the MLSys Kernel Contest, with a companion repo for evaluation and reproduction.

Key highlights

• Task contracts up front — objective, constraints, validation command, promotion criteria — so the agent knows what “done and fast” means before writing a line.
• Two pinned, MIT-licensed skills: KernelWiki and ncu-report-skill (the name suggests Nsight Compute reports; the README doesn’t elaborate).
• A Community Kernel Wishlist: optimization requests arrive as pull requests with a reproducible definition, representative workloads, and a best-known baseline — B200/B300 GPUs only for now.
• Harness-agnostic by design; downstream tasks bring their own evaluator, datasets, and profiling tools.
• Split licensing: CC BY 4.0 for docs and prompts, Apache 2.0 for first-party code, separate terms for the submodules.

Caveats

• Self-described early research prototype, still under active development.
• Claude Code-centric as shipped; porting the workflow to another agent is on you.
• The quickstart points at a mit-han-lab repository rather than this one, and KernelWiki’s upstream carries artifact snapshots under additional terms — the pinned submodule in this repo exists partly to omit them.

Verdict — If you’re pointing coding agents at GPU performance work, this is a rare mix of process discipline and contest-proven lineage — worth reading even if you never run it. If you wanted a turnkey kernel library or a benchmark harness, keep moving; this is prompts, docs, and a very good filing system.