From bare loop to self-rewriting harness: 21 papers map the history of agent harness engineering
omarsar0 · x · 2026-09-12
Elvis Saravia of DAIR.AI published a curated collection of 21 papers on Harness Engineering, based on YC Paper Club's Harness Edition. His core argument: build and own your harness instead of enduring out-of-the-box agent frameworks — even a minimal harness can quickly yield better code, outputs, costs, and writing.
- Definition: the harness is everything between model weights and the world — the loop, context assembly, reachable tools and skills, spawnable sub-agents, and increasingly the harness's own code
- Key insight: the same weight file can score 30% or 95% on the same benchmark depending solely on what surrounds it
- Evolution: from GPT-2's bare while-not-EOS sampling loop (2019), to few-shot context as the first lever (GPT-3, 2020), to harnesses that rewrite themselves by 2026 — six years legible as one repeated move of giving the loop new capabilities
Related event: 21-paper collection traces Agent Harness engineering history(2 posts)→
More from coding & agent
- HABIT-Bench: Best agent memory system scores just 33.61% at inferring user habits — SJTU · 2026-10-28
- Memory Tax study: dedicated agent memory systems often lose to naive RAG — UIUC-CS · 2026-10-28
- Regehr's 'Piano Test': would you trust formally verified code with your life? — moyix · 2026-10-28
- Vibe-coded 'AI 浮生记' game satirizes AI startup life with compute black-market trading — vista8 · 2026-10-28
- Cursor ships Annotate in design mode: agents take visual prompts via on-page markup — soleio · 2026-10-28
- The legendary fake Go generics: Unicode Canadian Syllabics as angle brackets — rickasaurus · 2026-10-28