AI is strong at implementation and weak at architectural judgment with no objectively right answer¶
Insight: Lalit Maganti spent eight years thinking about a library and three months building it with AI. His report: when intent was clear, AI accelerated execution dramatically; when intent was unclear, it sat 'somewhere between unhelpful and harmful', because cheap refactoring let him defer the architectural choices that actually needed making. Rebuilding once with more deliberate human design produced a more robust result.
Detail: Corroborates existing entries in this file about agentic defect accumulation and the architectural-awareness prerequisite for vibe coding: easy refactors are not free if they let you avoid deciding what the system should be.