Understanding is the new bottleneck

Geoffrey Litt’s argument turns on a distinction worth keeping. When agents write most of the code, the reason humans still need to understand it is not verification, checking the work, but participation: staying a creative collaborator on a project that continues to change.

That reframing changes what tooling you want. Instead of accepting cognitive debt as the price, he borrows three ideas from education. Explanations, meaning code explainer documents that build intuition and context before showing a diff, with quizzes to confirm the understanding actually landed. Micro-worlds, interactive environments where you can poke at a system and develop a feel for it, which agents can build for you. And shared spaces where a team converges on one mental model rather than several private ones.

He puts it in the older tradition of computing as augmentation rather than automation, dynamic simulation as a way of thinking better. The practical claim is smaller and more testable: getting deeper in the loop does not happen by default, and needs tools designed for it. Sits with AI agents and the refactoring that never happens, which describes what erodes when nobody keeps up.