Why Coding Agents Need Proof Obligations, Not Just Better Prompts

A developer built a complete application screen using six specialized AI agents in roughly two hours, but argues that code generation speed is no longer the real challenge — verification is. Instead of writing longer prompts, he assigned each agent a binary, independently checkable proof obligation: visual match with the design, four browser-tested user flows, a green test suite, and an empty punch-list. A dedicated design inspector compared the app's actual rendering against the reference mockup in both light and dark themes, while end-to-end flows were run in a real browser rather than simulated. The 167-test suite, including PostgreSQL store tests gated behind an environment variable, served as a final exit gate. Every mistake the agents made during the process was converted into a written harness rule, not a revised prompt, ensuring future runs could not repeat the same errors.