AI coding agents can produce software faster than humans can review it. Yet humans remain essential: only they know what they actually want. That makes human understanding the new bottleneck in agentic coding.

Olivier Gourment will share how ATDD helped him deliver a three-year, multimillion-dollar project in five months, then demo `acceptance_harness`: the Phoenix toolkit he uses today to preserve intent, generate reviewable evidence, and keep his codebases (and himself) sane.