Comprehension debt compounds silently
When something one-shots and seems to work, the temptation to move on without understanding it is strong. The debt accumulates across sessions until you are supervising an agent whose output you can no longer evaluate. The paradox: effective supervision requires the very coding skills that atrophy from overuse. Mitigate by staying engaged in implementation and reviewing every diff.
The mechanism is straightforward. Each time you accept output without fully understanding it, you carry forward a gap in your mental model. Over weeks, those gaps accumulate until the codebase exceeds your comprehension. You become dependent on the agent to explain its own work, and the agent has no memory of why it chose a particular approach three months ago. The defense is discipline: read every diff, ask the agent to explain anything unclear, and periodically code without AI to keep your discrimination sharp.
References
- Andrej Karpathy. Karpathy notes on Claude coding workflows and leverage
- Lars Faye. Agentic Coding is a Trap