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

影片總結

YC Harness Night shows Prime Agent reaching 95% on ARC-AGI, OpenJarvis cutting costs 800x, and QM scaling agents across YC.

At YC Harness Night, speakers argued that the harness—not merely model weights—determines what AI agents can accomplish. The same Claude Opus model reportedly rose from 30% to 95% on ARC-AGI with a stronger harness, while Prime Agent demonstrated persistent sub-agents, programmable memory, context management, and long-horizon work, including 633 agents producing 23 million tokens over a seven-day factorial run. OpenJarvis presented a privacy-focused local AI stack that uses cloud models to optimize on-device configurations, claiming up to 800x lower inference cost. YC’s QM showed how one open-source harness gives every employee personalized Slack and web assistants with sandboxes, scheduled jobs, internal data access, and shared resources—while highlighting unresolved challenges around persistence, permissions, social context, and agents giving up too early.

章節

1. 0:00Why harnesses matter: ARC-AGI rises from 30% to 95%
2. 4:27Building an auto-researcher by accident: eight H100 nodes producing papers
3. 7:13A five-minute history of harnesses: from GPT-2 loops to recursive RLMs
4. 13:56Self-improving harnesses: DSPy prompts to Darwin Machines editing code
5. 17:22Tonight's speakers: Seth Karten, OpenJarvis, and QM
6. 18:35Seth Karten: Prime Agent as a self-improving RLM harness
7. 21:44Context as an L1, L2, L3 cache: compaction, RAM, and refinement
8. 24:51From Turing machine to von Neumann computer: expressive harnesses and persistent sub-agents
9. 28:24Messaging between agents: coordination for long-horizon performance
10. 30:04ARC-AGI results: Prime Agent reaches 95.5% with Opus
11. 33:09Emulator Bench and GPU kernels: out-of-loop experiments scale to 633 agents
12. 39:21The five primitives of a personal AI stack: OpenJarvis runs locally
13. 42:47Letting cloud models optimize your local stack: OpenJarvis beats out-of-box deployments
14. 43:53800x cheaper than the cloud: local inference benefits from cloud optimization
15. 45:58Josh France and Regan Bell: QM, YC's agent harness for work
16. 47:29A history of YC's internal agents: from the January 2025 general agent to coding bots
17. 49:24OpenClaw and a fleet of 50 agents: YC scales personal assistants
18. 51:04Pulling the brain out of the sandbox: QM centralizes context in Postgres
19. 54:43Letting the agent choose its own sandbox and model: QM keeps the harness thin
20. 57:16The grind tool: budgets on goals prevent agents from quitting early
21. 58:50Agents don't understand social context: permission systems limit shared knowledge