Designing Self-Improving AI Agent Harnesses Around Human Priorities

I explain how to design and refine the systems surrounding coding agents so their work reflects changing goals, constraints, and definitions of cost. A steel-industry project that converts 2D technical drawings into 3D CAD provides a practical example; the focus is on the system that built the application, not a detailed product walkthrough.

- Engineering example: reconstructing 3D models from PDF drawings
- Harnesses and evaluation loops: adjustable instructions, tools, feedback, and cost per accepted change
- Iteration and learning: comparing harness variations and retaining useful experience through memory and policies
- Human priorities: accounting for time, money, urgency, and the value of exploration

This talk is aimed at people building or using AI coding agents and automated workflows. I outline ways to assess agent systems and suggest actions such as tracking costs, testing competing configurations, and preserving methods that may help on future tasks.