Back to Y Combinator

Self-Improving Harnesses, Local Personal AI And YC's Agent For Work | YC Paper Club

Y CombinatorSeptember 7, 20261h 0m
Topics43
YC Harness Club Introduction0:00The Value of Harnesses1:00Static Harness Era vs Self-Improving Harnesses2:00ARC-AGI and Harness Performance3:00Research Agent Swarm Implementation4:30Autonomous Research System6:00Historical Evolution of Harnesses7:00Context and Output Innovations8:30Multi-Agent and Reflection Systems10:30Harness V1 Architecture13:00Self-Improving Harnesses14:00Continual Harness and Memory Systems16:30Event Presenters17:30Prime Agent Architecture18:35Agent Session Structure19:30Persistent State and Memory Hierarchy21:00Context Management Strategies22:30Agentic Operating Systems and Harnesses24:34Persistent Sub-Agents and Session Management26:01Continual Harness and Learning Mechanisms27:02Inter-Agent Messaging and Long Horizon Performance28:31ARC-AGI Evaluation Results30:30Emulator Bench and Long Horizon Experiments34:00Multi-Day Factorial Runs and Technology Progression36:31Harness Design Recommendations37:00Local Personal AI with Open Jarvis37:31Open Jarvis Architecture and Primitives39:30Optimization and Cloud-Local Hybrid Approaches42:01Optimization Results and Future Trends44:30YC's QM Agent Harness for Work46:01Evolution of YC's Internal Agent Projects47:21Expanding Agent Capabilities Through Integration48:00OpenClaw Adoption by YC Partners49:30Scaling to Fleet Management with Hermes Agents50:00Unhobling Framework and Model Capability Expansion51:00Architectural Shift: Centralized Brain with Postgres52:00Automated Improvement and Human-in-the-Loop Systems53:02Resource Integration and Permission Systems54:01Dynamic Sandbox Allocation and Runtime Selection55:30Minimal Harness Architecture57:00Addressing Agent Limitations57:31Multiplayer Context and Social Awareness Challenges58:30Open Source Release and Hiring59:30
In a Nutshell

Harnesses now drive most AI progress through self-improving architectures that enable persistent memory, programmatic sub-agents, and test-time learning—outperforming static prompt engineering. Prime Agent, Open Jarvis, and YC's QM demonstrate this with concrete results: 95%+ on ARC-AGI, 800x cost reduction for local AI, and fleet-scale agent deployment across 50+ VMs. The key shift is from fixed tool loops to agentic operating systems where models dynamically manage context, spawn sub-agents, and accumulate capabilities through continual refinement.

AI-Generated Notes

These notes were generated by AI and may contain inaccuracies.

The event opened with a discussion of the new YC Paper Club visual design created by Ev, head of design at YC. The speaker explained that harnesses represent scaffolding and prompt engineering rather than traditional research, despite recent community pushback against considering prompt engineering as legitimate research at top-tier machine learning conferences.

Harnesses deliver an 18% performance improvement between different implementations. This difference determines whether ARC-AGI functions or fails entirely. The speaker referenced meter plots showing release dates versus agent runtime duration, demonstrating that much recent progress stems from harness improvements rather than model intelligence gains alone.

The speaker distinguished between the static harness era, where harnesses remain fixed, and the recent six-month period focused on self-improving harnesses. A plot from Trajectory's CEO illustrated how research emphasizes model perplexity and IQ metrics while neglecting test-time experience adaptation. The speaker described experiments increasing sample counts online and the challenge of learning from batch size one, noting that in-context learning saturates after 40-50 examples without further improvement on validation sets.

ARC-AGI emphasizes rapid adaptation to new problem distributions. Claude Opus achieved 30% on the private holdout set verified by Greg and Chalet. Harness improvements elevated performance to 95%, while AVO from Nvidia reached 100%. The speaker mentioned forking Karpathy's auto-researcher in March and accidentally building a harness while creating a user interface.

Sign in to read the full notes

Get access to AI-generated notes, topic timestamps, and more.