YC Harnesses Deep Dive: Same Weights Jump From 30% to 95% on ARC-AGI
ycombinator · x · 2026-09-07
YC convened frontier researchers and founders for a deep dive into agent harnesses, pushing back on the view that harnesses are 'just prompt engineering': the same model weights scoring 30% on ARC-AGI hit 95% with a better harness.
Key topics:
- History and self-improvement: a five-minute history of harnesses and self-improving designs; Seth Karten demos Prime Agent, a self-improving RLM harness.
- Context as cache: treating context as L1/L2/L3 cache, the Turing machine to von Neumann analogy, and inter-agent messaging.
- Benchmarks: ARC-AGI results under different harnesses plus Emulator Bench and GPU kernel generation.
- OpenJarvis local AI: Jon Saad Falcon's five primitives of a personal AI stack, cloud models optimizing the local stack, claimed 800x cheaper than the cloud.
- YC's internal agents (QM): building an agent for every employee, a fleet of 50 agents via OpenClaw, letting agents pick their own sandbox and model, the 'grind tool' for goal budgets, and lessons on agents missing social context.
More from coding & agent
- Agent Field Kit launches $29 security runbooks for production AI agents — alvelda · 2026-10-27
- Dev mixes Claude specs with abliterated GLM to build SS7 threat detector — l4rz · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- GitHub Next's Chopin prototype explores real-time multiplayer agentic planning — max__drake · 2026-10-27
- TMLR paper: enforced thinking doesn't reliably make LLM agents disclose key info to users — SharonYixuanLi · 2026-10-27
- LLM-written Rust library bbtree verified: speed claims hold, memory savings reverse — lmoroney · 2026-10-27