'Cognitive Debt' Is the Best Name for the Problem I've Been Feeling But Couldn't Articulate

Mario Zechner wrote an essay. Simon Willison amplified it. The core argument: agentic code generation creates "cognitive debt," where mistakes compound faster than humans can review them, and the speed that makes agents attractive is precisely what makes their failures catastrophic.

Zechner isn't some random critic. He created the Pi agent framework that powers OpenClaw. He's built the thing he's criticizing. That gives his argument real weight.

I've felt this in my own work. I run Claude Code in parallel sessions across multiple git worktrees. The velocity is genuinely 5-10x what I could do alone. But I've also shipped bugs that I wouldn't have shipped six months ago because I was reviewing agent-generated code at the same pace I was generating it. Which is to say, too fast.

The argument isn't that agents are bad. The argument is that speed without proportional review capacity creates a specific kind of technical debt that's harder to detect than normal debt. Normal tech debt, you can see it. You know the code is messy. Cognitive debt is invisible because the code looks clean. It often IS clean, line by line. But the architectural decisions, the edge cases not considered, the assumptions baked into generated patterns. Those compound.

Willison's framing is the one that stuck with me: "discipline to find a new balance of speed versus mental thoroughness." The bottleneck used to be typing code. Now it's understanding code. And understanding takes time that velocity-obsessed workflows don't budget for.

My practical response: I've started limiting the amount of agent-generated code I merge per day to what I can actually read and understand. Not skim. Read. Some days that means I generate 2,000 lines and merge 400. The unused 1,600 lines feel wasteful, but cognitive debt feels worse when it shows up in production at 3am.

For builders running autonomous agent loops: measure your review capacity honestly. If you can review 500 lines per hour with real comprehension, that's your throughput ceiling regardless of how fast your agents can generate. The agent's speed is irrelevant if your understanding can't keep up.