AI Might Not Take Your Job. But It Could Cost You Your Next Promotion.

Six days. A budget of a few thousand dollars. Machines that do not sleep, do not lose interest and do not need a cup of tea.

In July 2026, two dozen researchers led by Peter Kirgis and Sayash Kapoor at Princeton gave frontier AI agents a job instead of a task. They took the main research questions from two unpublished NeurIPS 2026 papers, handed them to Anthropic's Claude Opus 4.8 running on an open-source scaffold called OpenClaw, and said, in effect, go and do this work.

The agents did a great deal of it. They set up the technical environment. They managed the machines. They wrote code, ran experiments and worked through the night, without a single complaint. One detail from those six days bothers me. One agent declared the project complete seven hours before the deadline, shortly after its own reviewer had rejected the work again.

Then the original researchers read the two papers the agents had produced, and they rejected both.

Five problems came up again and again:

The agents had no sense of what good enough looked like.
When a research design failed, they pushed harder on the same design instead of trying another one.
They could not walk back out of a dead end.
They lost track of their own time and money.
And they slowly drifted away from the instructions they had accepted at the start.

Since the study came out, I have watched two comfortable conclusions form around it. I do not accept either one.

The first is relief. See, the machine cannot do our work after all. That is the lazy answer, and it will last exactly as long as the next model release. These agents were not weak. They ran six days of research on their own, without help.

The second is dismissal. Two research questions, a small number of runs, one model, one field. The researchers say this plainly in the paper, and they are right to. A study this size cannot tell you where the boundary will sit in two years, and it tells you nothing directly about accountants, designers or teachers, because it never looked at them.

But it can tell you something true about today. You do not need to drink the whole glass to know that the milk has turned. And what it shows is simple.

The agents were strong at the tasks. They were weak at the job.

A task is work somebody has already decided is worth doing, with a way to check whether it is done. Write the code. Run the experiment. Produce the draft.

A job is everything around those tasks. Choosing what matters, knowing when the answer is not good enough, throwing away four days of work because the approach was wrong, deciding when to stop.

Stop asking about jobs. Look at the work inside yours

It helps to put work on a line, from the easiest thing for a machine to the hardest.

At the near end sits work that is bounded and checkable. Find the information. Sort it and clean it up. Follow the steps in the process. Somebody has already decided that this work is worth doing, and anybody can see whether it has been done.

Move along the line and the work gets harder to check. Combine information from different places and find the pattern. Then it gets harder still. Weigh two options when neither one is clearly better. Advise a person who has to live with your advice. Decide what is worth doing at all, and when to stop.

AI is moving along this line, from the near end towards the far end, and it is moving fast. It already does the first few steps better than most of us. It is becoming good at the middle. At the far end, as those six days showed, it still gets lost.

So the honest question has nothing to do with your job title. Your title answers nothing.

Two people with the same designation, the same salary and the same desk can sit at opposite ends of that line.

The question is how much of your workweek sits at the near end.

The part that should worry you: where your judgement came from

Think about how you learned to tell good work from bad work in whatever it is you do.

A consultant learned which three slides mattered by making forty of them and watching a partner throw away thirty-seven.

An engineer developed a nose for where systems break by fixing hundreds of small errors.

A lawyer learned which precedent carries weight by reading a great many that did not.

A shift manager learned to see trouble coming on a factory floor by walking it every morning for six years, including the mornings when nothing happened.

All of that sits at the near end of the line. Retrieving, sorting, cleaning, following the steps, doing the same thing again and again until something in you started to notice.

It looks like waste. That is exactly why it is being automated first.

We are removing the ground where judgement grows, while keeping judgement as the job description.

No appraisal will mention this. Your output has gone up. Your manager is pleased. But promotions have never really been given for output, whatever the form says. They are given to the person whose judgement other people trust, and that trust is built on years of small calls that turned out to be right.

Now the output is nearly free, and the small calls are being handed to a machine. The loss appears three or four years later, in a room where somebody asks what you think, and you find that you have opinions but nothing underneath them.

Three habits you can adopt right now to develop Judgement

Keep some work slow, on purpose. Pick one task a week that you could hand to AI, and do it yourself instead. Not the dull parts. The part where you have to sit with a problem, get it wrong, and work out why.
Get into the rooms where decisions are argued, not where they are announced. If you are not invited, ask. Judgement is learned by watching experienced people disagree, and then seeing which of them turned out to be right.
Write down what you expect before you know the answer. What you decided, what you thought would happen, and what would tell you that you were wrong. Read it again six months later. Nothing has taught me more than my own old notes.

If you run a team, this is bigger than a personal problem. You are automating the exact work through which your next generation of judgement was going to grow, and nobody will send you a report about it.

Imagine two people doing the same job, with the same tools, producing the same report on the same morning. One of them can look at what has come out and say, this is confident and it is wrong. The other cannot. In a year, nobody will remember which of them typed faster. In five, only one of them will be running anything.

That is the last mile of thinking, and it is the only part that is still scarce.

The machine stopped seven hours early because nothing inside it could feel that the work was not ready. Somebody has to feel it. That feeling is built slowly, out of ordinary work that is quietly leaving your week.

Nobody is going to protect it for you. And nobody will tell you that it cost you the job you wanted.

Read More:

Peter Kirgis, Sayash Kapoor and colleagues, "Can AI agents conduct open-ended AI research? Early evidence from two case studies", July 2026.

Sayash Kapoor and Arvind Narayanan, "AI agents can't yet do open-ended AI research", 5 August 2026.

Michelle Kim, "AI's recursive self-improvement might not come so quickly after all", MIT Technology Review, 18 August 2026.