From Workarounds to Root Causes: Experience Using Agentic Workflows to Debug Complex(…)
summary
Porting llama.cpp to the browser via WebGPU exposes a bug chain across WGSL, browser runtimes, compiler rewrites, and Metal; a context-managed, re-steerable agentic workflow runs scoped probes (WGSL, Playwright-driven browser, Metal), uses file-based records and reviewer prompts, and moves from a found bit-equivalent workaround in ~2 h active agent time to a 12 h shape-and-size trigger matrix and a 47-min Metal-compiler session produced a byte-identical intermediate-output comparison that yields a loop-unroller root cause.
AI-written summary · about
chapters (14)
- 0:00Agentic workflows for debugging
- 1:01Team and project overview
- 3:08LLMs on edge devices demo
- 4:35Debugging complex GPU stacks
- 6:19Traditional debugging vs agents
- 8:24Interactive agent mode attempt
- 9:46Problem decomposition with probes
- 12:36Agent memory management
- 13:38Workaround for GPU bug
- 14:40Minimal reproduction
- 15:24Compiler optimization bug
- 16:49Probe-based debugging efficiency
- 17:10Research results and impact
- 17:52Q&A
related talks
34mPanel Discussion: Browsers & Wasm
01h 16mAnthropic Workshop: Build Agents That Run for Hours
29mOur web development workflow is completely broken
38mAgentic workflows that compound
31mA Dummy's Guide to Agentic Exploit Generation
27mAttacking the WebAssembly Compiler of WebKit
43mReal World Playwright Case Studies That Will Change Your Perspective!
31mLet the Main Thread Breathe!