Developer finds AI coding agent exposes gaps in his own specs, not its execution

A software developer using an AI build agent called Build discovered that the tool executes plans in strict, sequential phases and stops after each one awaiting human review, unlike most AI tools that over-generate code autonomously. He used it to refactor a fragmented notification system in a frontend codebase, consolidating scattered logic under a single scheduler across five planned phases. The process worked as intended until Phase 4 revealed a bug: the notification queue never reset between sessions because the original brief never defined queue lifetime or flush conditions. Build had implemented exactly what the specification said, and where the spec was silent, so was the agent. The developer concluded the flaw was not in the AI's execution but in his own incomplete planning, noting that the structured review process was what allowed him to catch and understand the oversight clearly.