Agents stall from local search, not lack of smarts: zero-score steps trap Claude Code at 0.44–0.51x
thomasahle · x · 2026-11-01
A reposted insight argues coding agents stall less from lack of smarts than from local search: each edit is judged by the immediate score it gets. Illustrated with GPU kernel optimization — every step from WMMA toward WGMMA scores 0 on tests, so an agent stayed on the WMMA peak at 0.44–0.51× of FlashInfer, unable to cross the failure valley toward the WGMMA peak that reaches 1.21× (0.74× at the same budget, which the human team crossed with more budget).
The lesson for agent engineering: evaluation mechanisms need to permit exploration through temporarily zero-score paths, or agents get stuck in local optima.
More from coding & agent
- Jez Humble's talk: making CI/CD work for fully agentic delivery — RealGeneKim · 2026-11-02
- Super-agent context design cuts task time to 12.4 min vs 20.4 min baseline — Hesamation · 2026-11-02
- Jez Humble: everyone is rediscovering trunk-based dev and real CI — RealGeneKim · 2026-11-02
- The AI plan looked perfect — so it built a fish tank — jimmykoppel · 2026-11-02
- A buffer-based workflow with Typefully and Claude Code for sharing AI work without burnout — brandon_galang · 2026-11-02
- Grok bot can now spin up webhooks for agents, killing automation plumbing drudgery — johnlindquist · 2026-11-02