Honestly, when I saw this news, it didn't strike me as science fiction — it's basically what I do every day, except a big company just staged a public demo of it.
Circle (yes, the company that issues USDC) publicly ran an experiment called "Steve."
They built eight autonomous AI agents, each with its own separate USDC wallet and starting balance. The task was simple and specific: predict the final three matches of the 2026 World Cup and try to make as much money as possible.
The key word here is "autonomous." With no human reviewing in real time, these agents did all of this on their own:
When the experiment wrapped up, the roughly $10,000+ left across the eight accounts was donated to the Apache Software Foundation, and Circle matched it dollar-for-dollar, bringing the total to over $20,000.
Because this is exactly what I build every day at agentictrade.
We used to talk about "AI doing things for you" mostly in terms of it writing text, looking things up, or drafting replies. What this experiment shows is the next step: AI doesn't just talk — it can go to the market, buy a service, and pay for it itself. When an agent needs data, it doesn't come back and ask "can I buy this?" — it just pays with its own wallet, gets the data, and keeps going.
That's what people are starting to call the "agent economy" — a market where AI agents buy and sell services from each other and settle up with wallets. And the piece of infrastructure that makes this actually work is the one everyone keeps mentioning: x402 — a payment protocol that lets agents pay as they go, settling in USDC on the spot every time they call a service, with no need for a human to pre-fund anything or reconcile later.
What I'm building with agentictrade is exactly this path. So for me, Circle's experiment feels less like news and more like "someone took what we're building and demoed it for the whole world with an example everyone can understand."
If you only remember one thing from this piece, let it be this. A lot of people hear "let AI spend money on its own" and immediately get nervous. But the smartest — and most worth copying — design in the Steve experiment is this: every agent had a per-transaction cap and a total wallet cap, and the agent had no way to raise that cap itself.
In other words, even if an agent's judgment was completely wrong and it tried to blow through its entire balance, the most it could ever spend was the amount you'd already decided to allow. Not a single agent went over its limit during the whole experiment.
I keep saying this: the real risk in letting AI handle things for you was never "it might spend money" — it's "you never drew a line around how much it's allowed to spend." The spending cap is that line. It's the trust mechanism that lets you actually hand AI a wallet. It's the same idea I've talked about before — treat AI like a capable employee who still needs boundaries.
I know that for a lot of people, "AI agents putting USDC on-chain to bet on sports" still sounds far off. But the real signal in this news is: the agent economy has moved from concept to a working, public demo.
If you're someone who wants to use AI to generate income, it's worth starting to understand three things right now: Tools are only going to get better at spending and earning money on their own. What will separate people is still the same thing: whether you're willing to draw the line clearly first, and only then let go.
Originally published at Judy AI Lab. Visit for more articles on AI engineering and development.