Why The Harness Matters More Than The Model | YC Paper Club
At a YC Paper Club night, researchers from Prime Intellect, Stanford and YC argue the harness around a model, not the model itself, drives big gains, covering self-improving harnesses, local AI and YC's QM agent.
The gist
-
The host says Claude Opus scored 30% on ARC-AGI's private test, while harnesses alone pushed results to 95%, and Nvidia's AVO reached 100%.
-
Seth says Prime Agent, built on a Python REPL and persistent sub-agents, scored 95.5% with Opus and ran 633 agents over a 7-day Factorio run.
-
John says local models trail frontier models by only 6 to 12 months, and OpenJarvis uses cloud models to tune local setups for 800x lower cost.
-
Josh says YC went from running Claude Code in VMs to a fleet of 50-plus Hermes agents, which proved hard to manage, before building QM.
-
Rean says QM keeps the harness thin, stores sessions in Postgres, sets time or token budgets so agents don't quit early, and needs strong permissions.
Y Combinator uses it
"By then a lot of our engineers started using Claude Code and Codex. And we realized that you could pretty easily run these in a VM. And then we hooked that up to a Slack tag, which was a pretty powerful medium for people who just wanted to like run a one-off code change."
▶ Play at 48:2591 creators recommend it, including Riley Brown, Liam Ottley, and Peter Yang.
See all on the Claude Code page
Y Combinator uses it
"We ended up provisioning a fleet of 50 plus Hermes agents that were running in VMs. These were definitely pretty helpful but they required a lot of configuring for people to get value out of them. It was just inherently kind of difficult to manage this fleet — a whack-a-mole s..."
▶ Play at 50:1212 creators recommend it, including Peter Yang, Alex Finn, and Garry Tan.
See all on the Hermes Agent page
Y Combinator uses it
"QM — there at YC we had this general agent that we use and they came out with QM a month ago and it is meaningfully better and is so functional and I use it every day."
▶ Play at 14:14What they recommend · 4 picks
"QM — there at YC we had this general agent that we use and they came out with QM a month ago and it is meaningfully better and is so functional and I use it every day."
"By then a lot of our engineers started using Claude Code and Codex. And we realized that you could pretty easily run these in a VM. And then we hooked that up to a Slack tag, wh..."
"In January of this year a lot of the partners started using OpenClaude and one thing to know about YC partners is that they're incredibly busy. OpenClaude was useful because it ..."
Also mentioned
Related episodes
Episodes covering similar topics
More from Y Combinator