We report on porting llama.cpp, a widely used LLM inference engine, to the browser via WebGPU. Failures in this setting are hard to explain: one wrong value can pass through shader code, browser runtime, compiler rewrites, native GPU tools, and GPU execution before becoming visible. llama.cpp’s debugging tools localize many problems, but the remaining cases can take days or weeks of expert effort across an unfamiliar stack.

We built a context-managed agentic workflow in which scoped agents use real browser and compiler tools, communicate through file-based records, and rely on reviewer prompts to decide whether evidence is sufficient. In our case study, a quantized byte-unpacking shader returned zeros in Chrome on Apple Silicon. The workflow found a bit-equivalent source workaround in ~2 h active agent time (~3 h wall-clock with one token-quota refresh). Because the workaround did not explain the platform bug, we changed the reviewer prompt to require mechanism-level evidence. A local-array shape-and-size trigger matrix then took ~12 h active across six sessions (~3.5 d wall-clock paced by daily quota refreshes), and a subsequent 47-min Metal-compiler session produced a byte-identical intermediate-output comparison, a standalone Metal reproducer, and a GPU loop-unroller root cause. The experience suggests that re-steerable workflows can shift complex browser/GPU debugging from weeks-scale manual digging toward lower-overhead monitoring, steering, and artifact validation.