This is a post about a specific feeling that I haven't seen named yet.
Debugging code you wrote is hard. Debugging code you wrote and forgot is harder. Debugging code that an AI agent wrote and you never read is a different category of painful.
There's no mental model to activate. There's no decision trail to trace. You're not trying to remember — you're trying to understand for the first time, in a broken state, under pressure.
This has been my experience shipping with AI coding tools for 6 months. The speed is real. The comprehension debt is also real. And the debt comes due in the debugging session.
The math behind it
AI agents write at roughly 100–200 lines per minute. A developer reading for genuine comprehension covers maybe 5 lines per minute. If you're watching an agent work, you're reading maybe 2–3% of what it produces before you feel pressure to move on.
The other 97% ships without you.
What I've been doing about it
I started building Overseer — a tool that watches your AI agent and generates plain-English explanations of every file change in real time. Not to slow the agent down. To keep you informed while it moves fast.
The side effect: those plain-English explanations are much faster to absorb than raw code. I now understand what the agent built in each session without reading every line.
MVP is nearly done. But this post is really just asking: does this debugging experience resonate? Am I alone here?"
Top comments (0)