Hacker Wires Idle Strix Halo NPU into a Coding Agent, Beats Grep at Semantic Search
stereohype · reddit · 2026-10-08
Reddit user stereohype's Halogen project exposes AMD Strix Halo's idle XDNA2 NPU via OpenAI-style endpoints and wires four models into a coding agent: semantic codebase search (15/20 vs 9/20 for grep on keyword-poor queries), near-duplicate file scanning (45 pairs in 8.4s), fast binary decisions (120ms, 78% accuracy), and injection screening (42% recall, 0 false positives, shipped as fail-open). A placebo-controlled experiment showed the NPU only accelerates 'where to look' (95% of task time stays GPU-bound), cutting a keyword-poor task from 18.7 to 13.6 minutes. Halogen 0.17.1 reduced NPU latency 30x and decode 40%.
More from coding & agent
- Stas Bekman: top coding agents all fail at concurrent multi-tasking — orchestrator plus sub-agents is the fix — StasBekman · 2026-10-08
- Evidence launches: image search for AI agents retrieves source-backed charts in 125ms — RichardsonDx · 2026-10-08
- The ideal WordPress AI agent finds the conflicting CSS rule instead of stacking another !important — HaktanSuren · 2026-10-08
- The agent-first app checklist: seven surfaces from WebMCP to ChatGPT plugin, buildable in a day — seanwbren · 2026-10-08
- Researcher's agent workflow: label 100 examples, have the agent scale to 10k pseudo-labels — ducha_aiki · 2026-10-08
- One user's daily agent stack: 7 AI agents split by task, with OpenAI's Dots starting a business — omooretweets · 2026-10-08