"Plan mode is dead": Nuanced's creator says AI coding plans no longer pay off

Ayman Nadeem, a former senior engineer at GitHub, argues in a new essay that "plan mode", the planning step built into AI coding tools, has outlived its usefulness. Published on 24 September 2026, the essay draws on Nuanced, the desktop AI coding app he built around planning.

Two jobs, one of them expiring

Plan modes did two things, Nadeem writes: they gave an agent instructions precise enough to act on, and they helped a human understand what was being built. The first job is fading, because models now explore a codebase and make reasonable assumptions on their own. The second matters more than ever, but a plan document is the wrong abstraction for it.

Nuanced was built for that second problem. A model can write thousands of lines of code in minutes, so a wrong assumption hardens into files hidden beneath the surface of chat.

Why the approach failed

Nadeem lists four reasons. He had conflated planning with a plan. Models improved at understanding large codebases, so separate instructions mattered less. Nobody wants to read AI-generated text. And planning was separated from building in a way that split the process artificially.

Users were surprisingly indifferent to the spec, which carried more information without more clarity. Instead of scrapping the plan, the team built a "Spec Tour" through its important parts, adding another layer of text. The workflow ran in a straight line: chat, generate spec, review, revise, approve, implement, review code, so users had to finish thinking before they could start building.

The line between planning and executing is dissolving

Nadeem points to Codex, where the two are merging. The old cycle of plan, approve, execute made sense when a wrong turn was expensive. Agents that act autonomously, test their own work and revise after inspecting the result allow a different loop: understand, act, inspect, clarify, adjust, act again. Planning still happens inside it, it just no longer needs an artifact called "the plan".

He is also critical of splitting plan mode and build mode: users had to judge whether a task deserved planning and remember to switch modes, a call he thinks the AI should make from context it already has. Keeping that understanding current as a team moves from five agents to hundreds, he says, is unsolved.