When the AI Drives You Nuts, Take the Wheel
Field Notes from 200+ Semi-Autonomous Sprints — Part 10
After nine posts about making agents work better, the last lesson is the simplest: sometimes you should just do it yourself.
After nine posts about making agents work better — managing their context, verifying their output, scaling them up, measuring their decay — the last lesson is the simplest: sometimes you should just do it yourself.
Not often. Less than you think. But when the moment comes, recognizing it quickly saves more time than any prompt optimization in this series.
It's Almost Always UI/UX
After 200 sprints, the vast majority of times I've grabbed the wheel were visual issues. Not close — overwhelmingly.
Agents have no eyes. I've said it throughout this series, and it matters most here. They can read CSS. They can modify layout components. They can set spacing to exactly the value you asked for. What they cannot do is look at the result. The padding is 16px like you specified, but the card still looks cramped because the content changed and the visual weight shifted. The animation eases in over 300ms like the spec says, but it feels sluggish in context. The button is centered, but it looks off because the surrounding elements aren't visually balanced.
These are taste calls. They require eyes, not tokens. And the signal is unmistakable: you've described the same visual problem three different ways, the agent keeps producing technically-correct-but-visually-wrong output, and you're getting frustrated. That's it. That's the signal. Close the chat. Open the file. Fix it in two minutes. The three rounds of back-and-forth you just burned cost more time and tokens than the manual fix ever would have.
Don't Grab the Wheel Too Early
The temptation runs the other direction too. 'It'll be faster if I just do it' is almost always wrong for anything beyond the trivial. A complex refactor, a tedious migration, repetitive boilerplate across a dozen files — those are exactly where agents earn their keep, and pulling them off the task because you're impatient is the most expensive habit you can build.
The discipline is letting the pipeline work. Let the agent struggle a bit. Let it retry. Step in when you've genuinely hit a wall that more tokens won't solve — not when you're just antsy.
The Series in One Line
This whole series has been about one idea: know your tools. Know what context costs. Know what agents drift toward. Know what your gates catch and what they miss. Know when to scale and when to hold. Know what to measure and what to ignore.
And know when to just drive it yourself.
This is the final post in the series. Start from the beginning: The Agentic Maturity Model — Where Are You, and Where Are You Going?