Why The Harness Matters More Than The Model

Name: Why The Harness Matters More Than The Model
Uploaded: 2026-09-11T00:00:00.000Z
Duration: 3611 s
Channel: Y Combinator
Description: YC harness researchers show scaffolding delivers the biggest capability jumps: identical weights score 30% on ARC-AGI, then 95% with a better harness. Talks cover a self-improving RLM harness, context as a cache hierarchy, a local AI stack 800x cheaper than the cloud, and YC's 50-agent fleet.

Why The Harness Matters More Than The Model | YC Paper Club

YCY Combinator

Full transcript

English

Summary：YC harness researchers show scaffolding delivers the biggest capability jumps: identical weights score 30% on ARC-AGI, then 95% with a better harness. Talks cover a self-improving RLM harness, context as a cache hierarchy, a local AI stack 800x cheaper than the cloud, and YC's 50-agent fleet.

Watch on YouTube

Core points (3)

1Identical model weights jump from 30% to 95% on ARC-AGI purely from better harnessing.
2Treat context like an L1/L2/L3 cache hierarchy with explicit agent-side garbage collection.
3Self-improving harnesses that rewrite their own prompts and code are now a research frontier.