Paper argues "The End of Harness Engineering": a 4B model auto-rewrites agent harnesses
blaizedsouza · x · 2026-11-01
A paper titled "The End of Harness Engineering" argues hand-writing code around a model is finished as a skill. A human harness is one guess that ships and runs unscored for months; a loop-written harness is a self-grading file where a 4B model reads failures, proposes 8 edits, and only the best-scoring edit survives. The big model stays frozen while the file around it is rewritten each round, shrinking the human job to writing the scoring set: "write the exam before the code." The post also links a step-by-step builder's guide for building an agentic harness with Jev, with code and architecture diagrams.
More from coding & agent
- Pydantic AI Adds Native E2B Sandboxes: Coding Agents Stop Running on Your Laptop — dl_weekly · 2026-11-01
- Survey of 57 enterprise employees examines how AI coding assistants really help and fall short — burkov · 2026-11-01
- ccbs: open-source terminal dashboard tracks your Claude Code 5-hour and weekly limits live — jeayese · 2026-11-01
- AI Makes Bug-Finding Easy; Deploying Fixes Is the Real Bottleneck, Says Venables — philvenables · 2026-11-01
- 20+ AI agents keep engineering a hardware product autonomously over the weekend — dee_hw · 2026-11-01
- Claude Code mod swaps compaction for a transcript pointer, author calls compaction a 'collective hallucination' — secemp9 · 2026-11-01