With coding agents writing faster than anyone can read, the real bottleneck has moved from producing code to reviewing it, argues developer Vinicius Brasil, and that includes your own git diff after the agent finishes.
His rule is simple: reject AI code when you cannot explain the approach, even when it works. He often discards an entire session and starts over, and notes the difference between the failed first run and the better second one was not a smarter model but more time spent consolidating the problem himself, so he drives the agent instead of being driven by it.
It is a tidy counter to the idea that faster generation is pure speed-up. The cognitive cost of vouching for code you did not reason through does not vanish; it moves downstream to review, where it is harder to pay. Shipping a diff you cannot defend is borrowing against whoever has to touch it next.