Paper argues 'The End of Harness Engineering': loops rewrite agent harnesses, humans just write the exam
blaizedsouza · x · 2026-11-14
A paper titled 'The End of Harness Engineering' is making agent-framework builders nervous. Its argument fits in one line: hand-writing the code around the model is finished as a skill.
The argument
- A human-written harness is one guess about one task on one afternoon; it ships and runs for months unread and unscored.
- A loop-written harness is a file that grades itself: a small 4B model reads failures and proposes 8 edits, and only the edit scoring best on held-out questions survives.
- The big model stays frozen the whole time; each round rewrites the harness around it.
Implications
- The human job shrinks to one artifact: the scoring set — the exam the harness is graded on.
- Frameworks become disposable, and 'writing the exam' becomes the job.
- The quoted Claude Fable 5.2 harness guide (Jev engineering + 9 prompts) claims Anthropic ran 1,000 agents with the same model, and that one evening of setup lets Claude do the work of a small dev team.
More from coding & agent
- Agent drives Unreal Engine via MCP: auto lighting, rendering and self-critique loop — HankYeomans · 2026-11-14
- Running Claude agent sessions visualized as characters in an Oasis game world — LeviTurk · 2026-11-14
- Vibe-Coded Game Hits 35.7K Plays and 10.5K Players in First Week on Spawn — TAbrodi · 2026-11-14
- WordPress UTM QA tip: accept cookies and submit forms without refreshing — HaktanSuren · 2026-11-14
- Using Claude to Visualize Ideas in Real Time During Team Calls — Flomerboy · 2026-11-14
- Vercel Token Pass doubles value: $150 in open-weight model tokens for $30/month — cramforce · 2026-11-14