Agent Speed Creates Cognitive Debt

The clearest July 2 signal came from Geoffrey Litt’s framing, echoed by Simon Willison: understand to participate. Digg surfaced Geoffrey’s practice of having coding agents quiz him on changes and building small “micro-worlds” to understand what happened. Simon’s notes sharpen the point with the phrase cognitive debt. The practical shift is that agent speed can now outpace human comprehension unless teams build review and explanation into the loop.

Code that ships faster but leaves the team less able to reason about it is not durable leverage. The hidden cost shows up later in slower debugging, weaker review, fragile handoffs, and fear of touching agent-authored paths. Product builders need workflows that keep understanding compounding alongside output.

Add an understanding stage to agent-heavy coding work. Trigger: any change large enough that you would struggle to re-derive it from memory. Context: diff, touched systems, tests, and the agent’s rationale. Tools: agent-generated quiz questions, targeted walkthroughs, disposable repro apps or fixtures, and architecture notes. Verifier: a human can explain the change, predict failure modes, and modify it safely. Budget: cap change size before mandatory explanation. Artifacts: quiz answers, micro-world repro, and a short “why this works” note. Stop condition: the reviewer can participate in the next iteration without treating the diff as magic.

Take one agent-authored change this week and ask the agent to produce five questions a reviewer should answer before approving it. If the team cannot answer them quickly, the change is under-explained.