tl;dr: The author built Nuanced, a coding app centered on plan mode, but concluded the approach failed because models improved enough to make explicit specs unnecessary, AI-generated planning docs were painfully unreadable, and separating planning from building disrupted the natural interleaved workflow. Modern agents work better in a continuous "understand → act → inspect → adjust" loop rather than a waterfall "plan → approve → execute" pipeline. The unsolved problem isn't planning itself, but how humans maintain a coherent mental model of systems being changed by many parallel agents faster than they can review.