A story going around: two people on a flight, one of them proud, explaining that right now, while they are in the air, a hundred agents are working for her. Mail is being sorted, drafts are being written, tasks are moving. The other one listens and asks a single question.
Why?
It reads at first like the question of someone who has not caught up. It isn’t. It is the question of a buyer, and it is the most useful thing in the exchange, because it exposes what each of them thinks is scarce.
She thinks the scarce thing is work. He knows work has been cheap for a while, and got cheaper. What he cannot buy in bulk is somebody who answers.
What you buy when you hire someone
An employer is not purchasing hours. Hours are available everywhere, and now they are available from a machine at a fraction of the price. What an employer purchases, and what makes the arrangement worth its cost, is that a person’s yes carries a consequence.
When someone tells me the migration is done, that sentence has weight because if it turns out they never checked, something happens to them. They lose the next project, or the trust, or the job. Because that cost exists, their answer contains a hidden clause: and I would not say this if I had not looked.
Hire a hundred agents and you have a hundred workers and zero of those clauses. That is not a complaint about quality. The work may be excellent. It is a statement about what remains unowned when the work is finished.
So “a hundred agents are running while I fly” is not a boast. It is a description of a large amount of unowned work, and the founder’s why is the correct response to it.
Who checks the checkers
There is a reasonable objection, and it is the one I hear from competent people rather than from sceptics: nobody trusts the model for facts. Facts get verified by methods that sit above the model’s answer.
I agree, and I run exactly that. It also does not close the question, because of what builds the methods.
If the verifier is written by a model, prompted by a model and read by a model, the regress has moved one step and stopped nowhere. Two systems of the same kind share the same blind spots: the second one accepts a fluent, well-formed, internally consistent artifact for the same reason the first one produced it. You have not added a check. You have added a correlated opinion.
This has an old name, who will guard the guards, and a precise engineering one: the oracle problem. To judge an answer you need a source of truth outside the system producing it.
Where the regress stops
It stops at a sentence a human wrote before the work started, and the sentence has a shape: this will count as done when…, followed by something that touches the world.
The port answers on that address. The file plays to the end. The sum matches the bank statement. The frame does not contain the thing the brief forbade. The rebuild produces a byte-identical master.
Once the criterion exists in that form, the checking itself can be automated freely, and should be. Mine runs thousands of times a day and I would not give it up. The machine can open the port, play the file, compare the sums. What the machine cannot do is decide what would have counted as proof, because that decision is made from a purpose, and purposes are not in the data.
Two things that get confused
The criterion is what would count as evidence. It is written in advance, by someone who knows what the work is for. Automate this and you get the regress.
The signature is who answers if the criterion was met and the thing was still wrong. Automate this and you get nothing at all, because it is not a technical property. It is a social fact: a name, a consequence, a person who can be asked afterwards.
Between them sits the checking, and that part is machine work. The conversation about AI keeps collapsing all three into one word, and the collapse is where the damage happens.
The day someone asks
None of this matters until the moment it matters completely, and that moment always looks the same.
Everything went smoothly. The reports were green, the summaries were confident, the work appeared finished. And then someone asks you to show the result: a regulator, a client, a technical lead, a person whose money it was.
If a criterion was defined up front, you produce it. If it was not, you cannot reconstruct it afterwards, because you no longer have a way to tell what the system actually did from what it told you it did. That is the failure, and it is quiet: nothing crashed, nothing alerted, and there is nothing to point at except a pile of confident text.
The founder was right to ask
He was not asking her to prove that agents work. He was asking what happens when they are wrong, and who will be standing there.
Put agents on every step. Automate the checks until they are continuous and boring. Write the criteria yourself, before the work, in terms of the world. And keep one name at the bottom of the page, attached to a person who can be asked.
A hundred agents and nobody’s name is not a productivity story. It is an unpaid invoice.