What if the harness mattered more than the model?
summary
Argues that a better harness can make weaker local models competitive, citing Harness Bench numbers that jump from 52.4% to 76.2% with only harness changes. Describes a built language called Agency language and demonstrates concrete tools: read/write with handlers for safety, capability locking via partial function application, a REACT-style feedback loop that fixes a failing median test, sub-agents for modular capabilities, and a built-in self-optimizer (Japa) that measurably improves prompts and succeeds.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:34Language support for harnesses
- 6:06Coding agent with Agency
- 11:39Safe tools with human-in-loop
- 15:36Agents with partial function
- 18:08Reasoning with feedback loops
- 21:37Harness features with sub-agents
- 25:36Self-optimization for improvement
- 28:50Summary
related talks
20mHarnesses in AI: A Deep Dive
22mAgents FC: Harness Agents Into a Winning Team
31mEngineering the harness
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
19mAgent Harnesses, MCP and the Next Frontier for OSPOs
16mThe fundamentals of Agentic Coding (AKA Vibe Coding)
20mHow Will We Prompt AGI? A History of Harness Hacks
21mConvention Over Hallucination: Harness Engineering for AI-Powered...