This session presents harness engineering: a structured approach to AI-assisted development that surrounds the model with guardrails rather than relying on better instructions. I will cover the workflow I use in a production codebase, a research and planning phase with a mandatory human checkpoint, an implementation phase, and a separate evaluation phase where the agent grading the work is not the agent that produced it.

The core of the talk is evaluation design: how to write grading criteria, why they should be weighted toward where the model is careless rather than where you care most, and how hard thresholds prevent one good score from masking a failure. I will demonstrate the full loop live against a real feature.
Attendees will leave with a workflow they can adapt, and a concrete framework for keeping AI-generated code to the standard their codebase already holds.