Plan Mode Is Dead

Earlier this year, the author was convinced planning would become the most important part of building software with AI. That belief was strong enough to inspire Nuanced, a desktop coding app built around it. But the product’s failure ended up changing the argument: plan modes, as they exist now, are losing their usefulness.

The original appeal was simple. AI could generate code at a speed and scale humans couldn’t inspect comfortably, so the author wanted a way to keep a coherent mental model of a system while machines kept rewriting it. In practice, though, the process got messy fast. Users were not eager to read long, AI-generated specs, the model was often good enough to make decisions on its own, and the workflow split thinking from building in a way that felt forced.

Nuanced tried to solve that by turning planning into a persistent document and a distinct stage before implementation. Threads would collect questions, surface ambiguities, and produce a plan before code was generated. The idea was to support the whole chain from intent to implementation, review, and verification. But the author says the real mistake was treating the plan as an artifact instead of designing a process that improved human understanding.

The bigger shift is happening in the agents themselves. As models got better at exploring repositories, making reasonable assumptions, and revising their own work, the old human sequence of plan, approve, execute started to blur. The author now sees the loop more as understand, act, inspect, clarify, adjust, act again. That still leaves a hard problem: how to keep people oriented as more agents change more of a system at once, without burying them in another slab of AI-generated text.

And the source gets blunt about the scale problem. The understanding issue only gets harder as you move from five agents to hundreds. The answer, it argues, won’t be a bigger spec. It’ll be better ways to surface the few places where human attention actually matters.