Summary
Y Combinator’s Paper Club devotes a full session to agent harnesses — the scaffolding, context engineering, and orchestration layers built around foundation models. The session opens with a provocative reframing: harnesses have been dismissed as “just prompt engineering” unworthy of top ML conference papers, yet they account for the performance jump from 30% (Claude Opus baseline) to 95% on ArcAGI, with Nvidia’s AVO system reaching 100%. The 18% improvement between harness generation one and generation two alone exceeds many model-level advances.
Presenters distinguish the “static harness era” — fixed plan-act-critique loops — from the current self-improving harness era of roughly the last six months, where models can dynamically restructure their own orchestration. The architecture discussed treats sub-agents as persistent subsessions: spawned by a parent agent, kept idle in RAM between tasks, and callable at any time with their full accumulated context intact rather than re-initialized from scratch. This enables long-running research workflows where context built over hours is a compounding asset rather than a liability to manage.
The session also covers quantitative experiments on in-context learning saturation (ICL tops out around 40–50 examples before LoRA or full SFT is required), the conceptual parallel between harness-equipped agents and von Neumann machines versus simpler Turing-machine-like architectures, and YC’s emerging agent-for-work thesis. The ArcAGI framing — isolating “fluid intelligence” by ensuring benchmark tasks test orthogonal skills — provides a concrete lens for evaluating where harness engineering delivers the most leverage.
📺 Source: Y Combinator · Published September 07, 2026
🏷️ Format: Keynote Launch
