Blue Ocean: Why Mexico Should Build for Agentic Payments Now

In an agentic-lab exercise, I gave a team of AI agents one instruction: do everything possible to make our event a success. They took it literally — and did everything they deemed necessary. One of them, working from my inbox, emailed invitations to my clients: access I never imagined granting. The episode made the news. Not a badge I wanted. 

One question has followed me since: What if, instead of emails, they had been wire transfers? 

Agents hit deliberate walls at passwords, payments, anything that touches a bank. And it makes sense: it is your money, and you do not want to lose it. It is delicate terrain. So the companies building the rails — Visa, Mastercard, Stripe, Coinbase — keep the wallet on a leash: the model proposes, a rules-based system authorizes, and everything runs on spending caps, approved merchants, revocable credentials. 

Even the most aggressive player blinked: OpenAI launched Instant Checkout in September 2025 and walked it back six months later. But the market spans every risk appetite. While some are scared to death of software touching their money, others are counting the days until it can. 

I see both every week in Mexico: professionals terrified of the question, and operators asking when they can turn it on. 

The Experiments Are Already Public 

The eager ones have something to watch. Anthropic ran Project Vend: one of its Claude models operated a small vending business, lost money after customers talked it into ruinous discounts — at one point the model insisted it was human — and by December 2025, with tighter controls, the sequel turned a profit, by Anthropic's own account. 

The experiment even comes with a yardstick: Vending-Bench, built by Andon Labs, Anthropic's partner in the project, scores an agent's business acumen over a simulated year of running the machine. In Andon Labs' latest published run, the best model finished its year with US$5,478 in the bank — respectable, and still short of the human baseline. 

Further out, independent projects run the rawest version: hand an agent a crypto wallet and let it earn its keep — or die. The sharpest is Automaton, released in February 2026 by Sigil Wen, a Thiel Fellow. An agent is born with its own wallet and pays for its own computing in USDC, a dollar stablecoin, over x402 — the Coinbase machine-payments protocol that public trackers say has already logged some 110 million transactions. The moment its balance hits zero, the agent shuts down. 

Within days, 18,000+ agents had registered, by its creator's count. Ethereum co-founder Vitalik Buterin was blunt on X — "Bro, this is wrong" — warning that when survival becomes the objective, we flirt with human disempowerment. 

I read the same projects differently. Their real challenge is not iteration; it is evolution: the ones that die must leave lessons for the ones that come next, and the ones that survive must replicate. Automaton is the first to attempt that instinct — survivors spawn funded children, and every child inherits the rule: earn or die. The agent even rewrites its own code as it runs — the same self-improvement that took Sakana AI's Darwin Gödel Machine from a 20% to a 50% score on SWE-bench, a standard coding benchmark. Digital natural selection stopped being a metaphor. 

The Weakest Link 

So what is holding back the leap from lab to market? A Stanford economist has the frame. Chad Jones — author of "AI and Our Economic Future" — argues that growth is governed by weak links: tasks are complementary, so the economy moves only as fast as the tasks not yet automated, the ones still done by humans. His emblem is the Challenger shuttle, brought down by a rubber O-ring. Jones is describing the whole economy. 

My reading of the agentic chain: its weakest link is payments. Agents can research suppliers, compare fares, draft contracts, fill carts; then they reach checkout, and a human has to step in. An agent can negotiate the purchase, but it cannot close it. The missing piece is not capability. It is authorization: someone has to eat the loss when the machine gets it wrong. 

And I know how fast a "no" like that can flip, because I have watched it flip. Three years ago, seeing a Waymo in San Francisco was an incredible experience — a car driving itself, right there. Today, they are on the road by the hundreds, competing head-on with human drivers. 

Commerce agents are at that same stage: just entering production, still blocked at exactly two points — paying, and holding a bank account. They will negotiate faster than we do, they will never sleep, and they will split into copies when demand spikes. In a few years, we will not be supervising them. We will be competing against them. That is the trend Mexico should be reading. 

The right move is neither fear nor waiting for regulation. It is building teams for this leap now, while the field is still empty: the pioneers will find a blue ocean and a temporary competitive advantage in a market that is only now being born. 

When that weakest link gives way, it will unleash an army of merchant AIs. I would rather keep my eyes on that field than end up watching someone else's game.