Comprehension debt and skill atrophy

The cognitive cost of agentic coding is the part most easily ignored and most likely to bite.

• Comprehension debt (the term surfaced in replies to agent leverage‘s source): when something one-shots and seems to work, it is tempting to move on without understanding it. The debt accumulates silently.
• Skill atrophy is real and measured. A practitioner essay cites an Anthropic study showing a 47% drop in debugging skills with aggressive AI use. Generation (writing code) and discrimination (reading code) are different skills; you can review code you would struggle to write.
• The paradox of supervision: effectively supervising the agent requires the very skills that atrophy from overuse. So heavy reliance erodes the capacity to rely on it well.

Mitigations from the sources: demote the agent to a secondary process, stay engaged in implementation, and never generate more than you can review in a sitting. See Claude Code for the workflow where this tension is sharpest.

Sources: 0007-karpathy-claude-workflow.md, 0008-agentic-coding-trap.md, 0012-addyosmani-llm-workflow.md

Edit wiki/comprehension-debt.md; see how we maintain the wiki.