p:: AI f:: Claude Code, Whiteboard
- Understanding is the new bottleneck — Geoffrey Litt (AI Engineer talk, Jul 2026)
The understanding gap that builds up when agents write code faster than you absorb it. Like tech debt, you can skip paying it for a while, but it comes due later. The term was popularized by Margaret-Anne Storey and Simon Willison.
Why understand at all
- To verify is the weak reason: it’s a thumbs-up/thumbs-down check, and agents keep getting better at checking their own work.
- To participate is the real reason. A project is many agent loops, not one. Your mental model of the system is what lets you come up with the next idea. Without it you can only nod along.
Technique 1: explanations + quizzes
Treat every finished agent task as a chance to produce an explainer, not just a diff. Litt’s /explain-diff skill outputs HTML or Notion pages with this structure:
| Layer | What it does |
|---|---|
| Background | Explain the existing system before what changed |
| Intuition | State the goal in plain words, explain the concepts, and optionally add an interactive figure |
| Literate diff | Walk through the changes as prose in a sensible order, with code snippets, instead of files in alphabetical order |
| Quiz | About five questions at the bottom |
- Rule: don’t ship (or approve a review) until you pass the quiz. Reading feels like understanding when it isn’t. The quiz forces the “do I actually get this?” check.
- The quiz works as a speed regulator, so the agent loop can’t run faster than you can understand it.
- Read the explainer first, then the raw diff.
Technique 2: micro-worlds
From Seymour Papert’s “Mathland”: you learn by living inside an environment, not by being told. Have the agent build you a small, playable tool for the system, then do the work yourself inside it:
- Step-through debugger for an interpreter: scrub through execution, inspect the stack and the rules that fired, and leave notes.
- Migration “command center”: instead of reviewing a one-shot port script you can’t judge, step through the port yourself with the old and new sites side by side.
The trade-off: the agent debugging for you saves time, but debugging it yourself in a tool the agent built is what builds understanding. Agent-written code that helps you understand other code is cheap now, so use it.
Technique 3: shared spaces
Teams need the same mental model to riff quickly. Put agent plans and explainers somewhere the team can comment, such as a shared doc, rather than in one person’s terminal.
Takeaway
Augment, don’t just automate (Alan Kay’s computer-as-learning-medium). Agents let you go deeper into the loop, not only step out of it.