What's this about?
Geoffrey Litt, a design engineer at Notion, argues it still matters that you understand the code your agents write — and that reading diffs line by line isn't the only way to get there. His talk offers alternatives: code explainer documents, quizzes that test whether you actually understood, and "micro-worlds" — small interactive environments for exploring how a system behaves.
What exactly is a "micro-world" here?
Instead of reading code to work out what it does, you build a small interactive sandbox where you can manipulate the system's inputs and watch its behavior directly. Think of the difference between reading a circuit diagram and having a bench with switches you can flip. For agent-written code — where the volume is too high to read closely — the argument is that comprehension can come from interacting with behavior rather than from reading every line.
What's the community saying?
- "I read the code." — Mitchell Hashimoto. The most-upvoted comment quoted this and extended it: I own the consequences of my code, and that's a responsibility an agent can't take.
- The throw-it-away strategy. A widely-liked practical alternative: let the agent write it, and if reading it feels unclear or needs a lot of explanation, discard it and start over or write it yourself — treating comprehensibility as an acceptance criterion.
- Agreement on the problem, dissent on the solutions. Several argued the techniques manage the volume without addressing the real risk: hallucinated explanations of complex code, and eventually an architecture too complex for the agent to extend.
- LLM-generated PR descriptions are actively disliked. One team reported theirs are universally unpopular — overly complex mechanical descriptions with no sense of motivation. And the deeper objection: if you're reading to check the LLM's work, an LLM's summary can't be the check.
- "Understanding has always been the bottleneck." A pointed dissent: LLMs speed up the easy part, typing, and are neutral or harmful on the hard part, which is understanding the problem and choosing how to solve it.