A staff engineer at Etsy argues that the agent harness matters more than the underlying model, especially for weaker or local open-source models. Drawing on HarnessBench research showing 20+ point performance differences from harness changes alone, the talk walks through building a coding agent using a new language called Agency. Seven progressive examples demonstrate key harness improvements: adding tools, safety via interrupts and handlers, partial function application for constrained autonomy, ReAct-style feedback loops, sub-agents for parallel task execution, and self-optimizing prompts. The core thesis is that investing in harness quality can reduce dependence on expensive proprietary models.
•32m watch time
49 Impressions