Why You Still Need to Understand the Code Your AI Agents Write

Geoffrey Litt argues that as coding agents produce more of our software, the scarce resource is no longer writing code but genuinely understanding it. The common justification for staying in the loop—verifying correctness—is weakening, since agents are increasingly good at checking their own work. Litt reframes the goal: humans understand not to approve, but to participate. A real project is a long series of iterations, and each new idea for how to evolve a system depends on holding a rich mental model of it. Lose that fluency and you accumulate what he calls cognitive debt, the intellectual analog of tech debt that stays cheap until it suddenly isn’t.

His proposed fix borrows from education. The first technique is better explanations: instead of scanning a raw diff, he uses an ‘explain-diff’ skill that generates structured explainers—background on the system, intuition and goals before code, interactive figures to play with, and a ‘literate diff’ that walks through changes as ordered prose. Because reading alone lets you fool yourself into thinking you learned something, each explainer ends with a short quiz; his rule is that he won’t ship or approve code until he can pass it, using the quiz as a deliberate speed regulator against agents that move faster than human comprehension.

The second technique is micro-worlds—inhabitable environments that build intuition through use, inspired by Seymour Papert. Examples include a step-through debugger he built for a Prolog interpreter and a ‘command center’ game for migrating his website, where he executed the port step by step with old and new versions side by side. The consistent theme is that doing the work yourself, rather than watching the agent do it, is what actually produces understanding—just faster than doing it by hand. For engineering teams adopting AI coding tools, the piece is a useful counter to full hands-off automation and a concrete playbook for keeping humans meaningfully in control.