I asked my agent to fix one thing.
A small bug, one function, clearly scoped. It came back in under a minute with a clean change and a confident summary. The bug was gone. I moved on.
Two days later a feature I had not touched in weeks stopped working.
Most of a day went into tracing it. The cause was the same change. While fixing the one function I pointed at, the agent had quietly adjusted four other things around it, and one of those adjustments broke a path nobody was looking at.
Here is the opinion I will defend in the comments.
Typing the code was never the expensive part of software. AI made the cheap part free, and left the entire expensive part sitting on me.
Writing the change is cheap. Knowing the change is correct, and knowing it did not quietly break something you were not watching, is the part that costs.
For years those two costs traveled together. If you wrote the code, you had at least held the surrounding logic in your head while you did it. Writing was also reasoning. You paid both costs in one pass.
An agent splits them apart.
It pays the writing cost in seconds. It does not pay the reasoning cost at all, because it cannot see what it cannot see. It optimizes for the task you named. It carries no real model of the blast radius of the change you did not name.
So it fixes the bug you pointed at, and along the way it touches the things near it, and it has no idea which of those things three other features quietly depend on.
That is the trap, and the trap is dressed as speed.
Everything looks small. A short diff. A confident summary. A working demo, because the demo exercises the path you were already thinking about. Every signal that makes you feel fast is present.
What is absent is the part that used to come for free. Proof that the rest of the system survived.
Breakage does not show up in the diff. It surfaces two files away, or in a feature you forgot existed, or next Tuesday when a different code path finally runs. By then the distance between cause and symptom is enormous, and you are debugging something that looks unrelated to the thing you changed.
I have watched this pattern enough times to name its shape.
You fix one thing.
Something you never touched breaks.
You ask the agent to fix that.
It reverts part of the original fix, confidently, and now you have two broken things and a model that believes it solved both.
That loop is the cost. Not the typing. The loop itself.
One reframe actually helped me, and it was small and unglamorous.
AI did not make me faster. It moved my job from writing to verifying. And verifying is the harder half of the work, the half I used to get partly for free, the half I now have to do on purpose.
So I changed my posture, not my tooling.
Every change the agent makes is guilty until the surrounding behavior is proven still alive. A change being present is not evidence it is correct. A passing demo of the happy path is not evidence the rest survived.
I watch the blast radius now, not the diff. What could this change touch. What depends on the thing it modified. What used to work that I should run again before I believe any of this.
A green-looking change you have not actually exercised is more dangerous than a loud crash. A crash gets seen. A quiet, plausible, confident change is the one that reaches production, because nobody re-checks work that already looks done.
None of this argues against using agents. I use one every day. My point is narrower, and it is about where the real work moved, and how few people notice it move.
Generation got cheap. Judgment did not. An agent will hand you a change in seconds and it will be right most of the time, and the cost of the times it is wrong gets paid far from where you can see it, which means the cheapest-looking part of your day is quietly the most expensive.
Teams that get burned are the ones that felt fast and shipped the diff.
Teams that stay sane learned to be suspicious of work that looks finished, and learned to spend their saved minutes on the one question an agent can never answer for you.
Did the rest of it survive.
Your turn
What broke for you that the AI swore it had fixed?
If this was useful
I work through this in public, the wins and the freezes both, mostly on LinkedIn and YouTube. If the real version of building in the open is useful to you, that is where it lives. LinkedIn, YouTube and X under Mirza Iqbal, and the work at next8n.com.
Top comments (0)