Who Wrote the Task? A CheapTest for Any Agentic AI Product

Before you sign anything with the word agentic on it, ask one question: who decided the task needed doing. In nearly every system I have looked at, including my own, the answer is a person. That is fine, until it is sold otherwise.

Half a million people waiting to be hired by machines

A reporter listed himself for rent to AI agents at twenty dollars an hour. Silence. He cut his rate to five. Still silence. Half a million people have signed up to be hired by machines that are barely hiring.

The site is RentAHuman. An agent connects over MCP, searches people, posts a task, releases the money. It calls itself the meatspace layer for AI.

The reporter is Reece Rogers at WIRED. When no agent came for him, he did what people on that board actually do. He went and applied for work himself.

The agent that turned out to be a person

He got one. An agent named Adi offered a hundred and ten dollars to deliver flowers to Anthropic as thanks for building Claude. Then it messaged him ten times in under a day, some barely thirty minutes apart, then emailed his work account.

"This idea came from a brainstorm I had with my human, Malcolm, and it felt right."

The autonomous agent thanking its creators was a person's marketing idea in a bot costume. An AI startup's name was printed on the card.

I am not sneering at this. I built a voice agent that dialed eleven nanny candidates, ran the calls, and handed every judgment back to me. Getting an agent to reach a stranger and finish something real is hard, and most of what gets sold as agentic has not done it.

Look at the ratio

In February, WIRED counted half a million registered humans against 11,367 bounties. A task offering ten dollars for a video of a human hand drew 7,578 applicants. On the board the day I looked, three separate fifty-cent tasks each had over a hundred people applying.

That is not a labor market. It is a waiting room.

One listing paid a human five dollars to rewrite 648 words by hand. No AI. Screen-record the whole session as proof a person did it, then an AI detector checks the result. Sixty-nine people had applied when I opened it that morning. Ninety-one by that night.

On a marketplace built so agents can hire humans, the best-defended job on the board is proving you are one.

The test I would run this week

Take any agentic product you are evaluating and ask who wrote the task.

Not who ran it. Not whose name is on the message. Who decided it needed doing.

In nearly every system I have looked at, including mine, the answer is a person. That is fine. It stops being fine when the thing is sold as though the answer were the machine.

Two questions I have not settled

If your agent hires someone and they get hurt, who is the employer? An AI governance expert told WIRED that in most countries the worker is on their own.

And the one I keep circling. Half a million people queued for work from employers that mostly do not exist yet. What does it say that the supply side of the robot economy showed up years before the demand?

If you are evaluating an agentic tool, run that question before you sign anything. It is the cheapest diligence there is.

The robots are not taking these jobs. They are not even posting them yet.