Which Day Are We On?
The shopper with no hands
Here is a small puzzle. Picture a computer program in the middle of a job. It needs something it doesn't have. In the old days, a person had to step in, open an account and click "buy."
Now nobody steps in. The program buys it and keeps going.
In August 2026, Circle described a marketplace with more than 900 paid services that software could buy on its own, paying from a balance of digital dollars. Most of these were services one program calls through an API, a doorway that lets two pieces of software work together.[1]
It sounds small. It isn't. A purchase that once needed a person is now part of the job. Follow the money one more step, and a stranger question appears.
Two roads
Meet two made-up people. Maya works at a help desk. Sam owns a small help-desk business. Both use the same smart software, and Maya is just as good with it. Sam's business takes in more than it costs to run, and he keeps the difference. Maya gets a paycheck.
Same tool. Same skill. A totally different claim on the money it makes.
In May 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents, the people who answer customer questions, after they got an AI assistant. On average, they solved 15 percent more problems per hour. Here is the surprise: beginners gained the most. The top workers sped up only a little, and their quality slipped a bit.[2]
But the study measured work done, not pay. When a worker improves, the employer decides what to do with the extra output: raise pay, cut prices or hire fewer people. The money just reaches Maya and Sam by different roads: the paycheck road and the owner's road.
Two clocks
Why now? Because the machines are improving fast, in two ways.
In January 2026, a research group called METR tested AI agents, programs that carry out a string of steps on their own. It asked how long a task, measured in human time, an AI could finish half the time. That 50 percent mark is a yardstick, not an error rate for every job. AIs ace most very short tasks and fail most very long ones,[6] so the halfway point is a handy line for comparing models. For the models METR re-tested from 2023 onward, that length doubled about every 131 days on its software and research tasks. If the trend held, four doublings would make it 16 times longer in 524 days.[3]
That is the first clock. The second is easy to miss.
In an earlier study, METR asked a stricter question: how long a task could an AI finish four times out of five? Those tasks were four to six times shorter. But their length was doubling at about the same pace.[6] The gains seemed to come mostly from AIs getting more reliable and better at fixing their own mistakes.[6] One clock tracks how big a job an AI can take on. The other tracks how often it gets that job right. Both are moving.
Jeremy Allaire, who runs Circle, has written The Agentic Economy Treatise. He pictures firms of AI agents, software that directs them and a few people, using profits to buy more agents.[4] Allaire has a business stake here, since his company works on these payments. That is also why he watches the link between software, payments and business ownership so closely.
Here is the catch. A worker who only earns a wage gets paid for the work. One who also owns shares gets a slice of the growth. Deals made while the company is small keep steering the money long after the software takes over more of the work.
Builders and owners
This needn't be gloomy. Cheap, smart software could help one person start a business that once took a team. Economists Daron Acemoglu and Pascual Restrepo point out that while automation replaces people in old tasks, new tasks can raise the demand for human work and its share of income.[5]
Ownership is another way in. Workers can hold equity, meaning a small piece of the company. But a slice is only worth what comes with it. Allaire suggests wider ownership, including through digital tokens, but notes that voting power can stay with the biggest owners.[4] A small holder may get almost no say. And if your savings and paycheck both ride on one company, one failure can wipe out both.
Back to Maya. She answers the same questions all day. But she knows what the steps leave out: where customers get stuck and which mistakes cause trouble later. Suppose her boss gives her time to build an assistant from the company's own guides. Maya checks what it gets wrong and fixes it. She is no longer just following steps. She is building them.
Now think about the two clocks again. Maya's new job is catching the machine's mistakes. It exists because the machine still makes them. If the second clock keeps running and the errors keep falling, there will be fewer mistakes to catch, and fewer people needed to catch them. With this dataset, METR couldn't confidently measure task lengths at success rates as high as 19 times out of 20.[6] But the direction is clear enough to be unsettling.
That is why it must start now, while people like Maya still have a paycheck and are still needed in the loop. They need paid time to practice. And anyone who helps build a better process should get a share in the value it creates. Otherwise, the reward for helping automate a job could be losing the income that made learning possible.
So, who owns the machine that does the work? The answer starts with whoever owns the business, and with whatever rights workers and others hold in it. Teaching people to build with AI gives more of them a chance to create those businesses. A stake ties that effort to the money the machines make. That way, helping build a more productive economy can also mean having a place in it.
Sources
[1] Circle: Building the Open Agentic Economy
[2] Brynjolfsson, Li and Raymond: Generative AI at Work, QJE, May 2025
[3] METR: Time Horizon 1.1, January 29, 2026
[4] Jeremy Allaire: The Agentic Economy Treatise
[5] Acemoglu and Restrepo: Automation and New Tasks, JEP, 2019
[6] Kwa et al. (METR): Measuring AI Ability to Complete Long Tasks, March 2025