As coding agents produce changes faster than people can absorb them, designer and developer Geoffrey Litt argues that human understanding has become a central constraint on software work. His position is not limited to checking whether an agent's output is correct. Understanding also gives people the concepts needed to propose the next change and participate creatively in an evolving system.
Litt describes software projects as repeated loops with an agent rather than a single delegated task. A person who loses track of architecture and behavior may still approve an individual result, but becomes less capable of deciding where the product should go next. He connects that loss of fluency with cognitive debt: short-term speed purchased by postponing comprehension.
The first technique he proposes is a structured explainer. His `/explain-diff` skill produces HTML, Markdown or Notion documents that teach the existing system before describing a change. An explainer states the goal, introduces relevant concepts and orders snippets as a narrative instead of presenting files alphabetically. Interactive diagrams can build intuition before a reviewer reaches implementation detail.
Reading alone can create a false sense of comprehension, so Litt adds short quizzes. His explainer packets end with five questions about the change, and he says he will not send code onward until he can answer them. The quiz acts as a speed regulator, forcing the human review loop to pause when generation has moved beyond the reviewer's actual grasp.
A second approach uses micro-worlds: interactive environments that expose a system's internal state. Litt describes building a debugger for a Prolog interpreter that allowed him to move through execution, see the stack and inspect which rules were being evaluated. The purpose was not only to help an agent find faults, but to give the human a manipulable model of the program.
Shared spaces complete the broader idea. Explanations and interactive artifacts can become collaborative objects where teammates discuss what changed and why. Litt notes that he works at Notion and uses its embedded interactive HTML feature, a potential source of bias in his tool choices.
The techniques impose additional work, and the talk does not claim they are the only way to review agent output. Their common principle is that explanation should be generated alongside code rather than reconstructed later. If agents accelerate implementation, the same systems can help produce educational material that keeps human judgment connected to architecture, intent and future design choices.
In Litt's formulation, the bottleneck is not the agent's ability to produce another revision, but the person's ability to retain enough of the system to direct that revision intelligently.
