← back
Understanding is the new bottleneck — Geoffrey Litt, Notion
Takeaway
Use AI to teach you the systems it changes so faster coding does not erode your ability to shape the project.
Summary
- As agents produce larger changes, human understanding becomes harder to maintain even when automated verification improves.
- Understanding supports creative participation and future decisions, beyond checking correctness.
- Cognitive debt accumulates when developers accept changes they cannot explain.
- Litt uses explainer documents, quizzes, and interactive microworlds; his explain-diff workflow teaches background and intuition before implementation details.
cognitive-debtcode-understandinginteractive-learning
Original description
Autonomous loops are hot, but the reality is that most agentic tasks still require human judgement. And to guide your agents well, it's not enough to just verify correctness -- you actually need to understand the work they're doing. In this talk, I'll share some techniques for staying in the loop and efficiently developing understanding, combining old ideas from education and cognitive science with modern agent capabilities. You'll walk away with some practical tips for moving faster with agents by understanding more, not less.