Are AI Agents Replacing Developers? What Running Them Daily Actually Changed

Two stories are competing for the same headline. In one, AI agents have made software engineers obsolete and the sensible move is to stop hiring. In the other, engineers are resigning over what the labs are building and the whole thing is a bubble about to correct.

I run coding agents every working day across a portfolio of iOS, web, and automation projects. The reality I see is narrower and duller than either version, and more useful if you are deciding where to spend money.

What actually got faster

The honest answer is: the part between knowing what you want and having a first version of it. That part collapsed. Scaffolding, boilerplate, test fixtures, migrations, the second and third variant of something that already exists — that work is close to free now, and pretending otherwise is expensive.

What did not get faster is deciding what to build, knowing when the output is wrong, and owning the consequence when it ships. Those were always the expensive parts. They still are, and they got proportionally more expensive, because the cheap part now produces volume that has to be reviewed.

This is the bit both stories miss. Agents did not remove the bottleneck; they moved it downstream to judgment and verification, where it is harder to see and easier to skip.

The failure I did not expect

I assumed the difficulty would be code quality. It was not. The difficulty was knowing what the agent had actually done.

Running several agents across several repositories, I could not reliably tell live work from yesterday's transcript, or a half-written record from a finished action. A dashboard that shows you a partial write as though it were complete is worse than no dashboard, because you will act on it.

That is why I ended up building NexusTerm — a native macOS terminal that rebuilds agent activity from the transcript files rather than tailing them live, and only ingests a record once it is complete. The engineering that mattered was not the AI part. It was the boring correctness work: byte boundaries in Unicode fixtures, file replacement mid-read, distinguishing stale activity from current activity.

That is the shape of most real agent work. The model is the easy dependency. The system around it is the job.

What this means if you are hiring

The question "do I still need a developer?" is the wrong one. Better questions:

Do you need someone to produce code, or to be accountable for it? Production is cheap now. Accountability is not, and it is not something you can buy from a model. Someone has to decide the thing is correct and carry it when it is not.

How will you tell good output from plausible output? Agents produce work that reads well and fails in specific conditions. If nobody on your side can evaluate it, you have not saved money, you have deferred the cost to whoever discovers the bug.

Is the work novel or is it the fourth variant of something known? For the fourth variant, agents are genuinely transformative and you should expect to pay for far less time than you did in 2023. For anything load-bearing, novel, or regulated, the estimate should not have moved much.

The bubble question

Some of the correction people are predicting will happen, and it will be concentrated in exactly the places where the claims outran the evidence — the "fully autonomous" pitches, the agent startups with no verification story, the deployments that were never measured against a baseline.

It will not touch the unglamorous version: bounded workflows, explicit permissions, a human approval step, and a number you agreed on beforehand. That version was never dependent on the hype and will not be damaged by its absence.

If you want the boring version, agentic automation is where I would start, on one workflow with a measurable baseline. If you are evaluating someone else to do it, the AI developer hiring checklist is the list of questions I would ask.

Neither story is going to resolve soon. Your next project does not need it to.