@songhan_mit: Explore KDA (Kernel Design Agents): https://github.com/mit-han-lab/kernel-design-agents…
Summary
KDA (Kernel Design Agents) is an open-source agent-centric workflow from MIT HAN Lab for using coding agents to research, implement, verify, and iterate on performance-sensitive CUDA kernel tasks. It helped Databricks achieve rank #1 on NVIDIA's SOL-ExecBench leaderboard.
View Cached Full Text
Cached at: 07/02/26, 04:18 AM
Explore KDA (Kernel Design Agents): https://github.com/mit-han-lab/kernel-design-agents…
mit-han-lab/kernel-design-agents
Source: https://github.com/mit-han-lab/kernel-design-agents
Kernel Design Agents
Kernel Design Agents (KDA) is a agent-centric workflow for using coding agents to research, implement, verify, and iterate on performance-sensitive CUDA kernel tasks.
This repository documents the early research prototype and is still under active development (we are looking for community feedbacks!). If you are interested in HAN Lab Mafia’s solution ranking #1~3 on tracks at MLSys Kernel Contest, please refer to mit-han-lab/mlsys2026-flashinfer-contest for perform evaluation and reproducement.
Contents
| Path | Purpose |
|---|---|
docs/agent-flow.md | Minimal end-to-end KDA workflow. |
prompts/README.md | How to use prompt templates. |
prompts/basic-flow.md | Generic starter prompt for a new task. |
CLAUDE.md | Repository-facing agent instructions. |
Getting Started
Install the agent workflow dependencies before starting the agent session:
git clone --recurse-submodules https://github.com/mit-han-lab/kernel-design-agents.git
cd kernel-design-agents
# link skills
mkdir -p ~/.claude/skills
ln -s "$(pwd)/skills/ncu-report-skill" ~/.claude/skills/ncu-report-skill
ln -s "$(pwd)/skills/KernelWiki" ~/.claude/skills/KernelWiki
# or clone skills directly
mkdir -p ~/.claude/skills && cd ~/.claude/skills
git clone https://github.com/mit-han-lab/ncu-report-skill.git
git clone https://github.com/mit-han-lab/KernelWiki.git
Install the humanize Claude Code plugin from the Claude Code plugin UI:
/plugin marketplace add PolyArch/humanize
/plugin install humanize@PolyArch
Minimal Flow
- Create a separate implementation workspace for the target task.
- Define the task contract: objective, constraints, validation command, and promotion criteria.
- Start an agent session in the implementation workspace.
- Give the agent
prompts/basic-flow.md, filled in with the task-specific details. - Ask the agent to write a short plan draft to
docs/draft.mdin the implementation workspace. - Convert the draft into an executable plan, either manually or with a planning tool such as Humanize.
- Implement in small iterations, verifying after each meaningful change.
- Record candidates, benchmark or evaluation results, profiling evidence, and final promotion decisions.
The workflow is intentionally independent of any single benchmark harness or hardware target. A downstream task can add its own evaluator, datasets, profiling tools, and domain-specific references.
Recommended Workspace Layout
Use this repository as reference material, then do implementation work elsewhere:
task-workspace/
docs/
draft.md
plan.md
runs/
outputs/
profile/
benchmark.csv
candidates.jsonl
The exact files can change by domain. The important rule is that the agent records enough context for another engineer to understand what was tried, what passed validation, and why the final candidate was selected.
Yuchen Jin (@Yuchenj_UW): Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉
What’s crazy is: we 100% leveraged AI agents to beat the competition.
This is a sneak peek at recursive self-improvement. The core
Similar Articles
@LigengZhu: Excited to share the KDA: Kernel Design Agents that powers HAN Lab Kernel Mafia top ranking #1~3 kernels at Kernel Cont…
KDA is an agent-driven kernel design framework that helped HAN Lab achieve top rankings in the MLSys FlashInfer Kernel Contest by minimizing human involvement. The agent leverages Humanize, KernelWiki, and profiler skills to produce state-of-the-art kernels.
AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents
AgentKernelArena is an open-source benchmark for evaluating AI coding agents on GPU kernel optimization, assessing full agent workflows and generalization to unseen configurations across 196 tasks.
@levidiamode: 163/365 of GPU Programming Looking at a few different agentic GPU kernel optimization systems today. The two I'm most i…
A tweet discussing two agentic GPU kernel optimization systems: Auto GPU Kernel by @dogacel0 and Kernel Design Agents from @songhan_mit's lab, both winners at the MLSys Sparse Attention FlashInfer competition. The thread highlights different approaches using subagents and Claude skills for GPU programming.
AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
Researchers from Carnegie Mellon, University of Washington, and Arm propose AdaExplore, an LLM agent framework for GPU kernel code generation that achieves 3.12× and 1.72× speedups on KernelBench Level-2 and Level-3 benchmarks through failure-driven adaptation and diversity-preserving search, without additional fine-tuning.
@songhan_mit: We develop an agent-native approach to accelerate genAI, continuing the success of KDA (Kernel Design Agent) at a highe…
Enze Xie announces Sol Video Inference Engine, an agent-native, training-free full-stack accelerator for video diffusion that auto-tunes cache, sparse attention, token pruning, quantization, and kernel fusion, achieving >2× end-to-end speedup on large models like 64B Cosmos3-Super and 22B LTX-2.3.