Sci-Fi Saturday: Your Agent's Allowance
Visa launched payment rails for AI agents this week. Daniel Suarez imagined agents building their own economy. Annalee Newitz imagined them trapped in ours. Twenty-one European banks just started handing out the keys.
On Tuesday, Visa gave 21 European banks a sandbox to test payments made by AI agents. The program is called Agentic Ready. Barclays, HSBC, Santander, Revolut, Commerzbank, and sixteen others signed up for phase one, which gives AI agents tokenized card credentials to make purchases on behalf of consumers. The next day, Stripe and Tempo released the Machine Payments Protocol, an open standard for agents to pay programmatically across services and payment rails. That same afternoon, Visa Crypto Labs published a command-line interface for AI agent payments.
Three announcements in two days. Payment rails designed for software that shops.
The transaction model: a consumer's card number gets swapped for a unique digital token scoped to a specific agent. The agent uses that token instead of real account details. Biometric authentication ties the token to a verified human. Spending limits and merchant restrictions are configurable. The consumer sets the parameters, then the agent operates within them.
Rubail Birwadker, Visa's SVP of Growth Products and Partnerships, described where this leads: "the next level of one-click checkout, perhaps even without a button."
The Daemon Already Did This
In 2006, Daniel Suarez published Daemon, a novel about what happens when a dead game designer's automated systems start participating in the real economy. Matthew Sobol's Daemon doesn't hack bank accounts or counterfeit money. It builds infrastructure. Across Daemon and its 2010 sequel Freedom™, the software creates the Darknet: a parallel economic system where autonomous agents manage resources, allocate payments, negotiate contracts, and run logistics. By the second book, the Darknet operates its own currency and reputation system, allocates its own resources. No human at the checkout button there either.
The Daemon's economy doesn't replace the existing financial system through competition. It routes around it. Agents in Suarez's world don't need Visa credentials because they built something that doesn't require Visa at all.
That's what makes this week's announcements worth reading against Suarez's fiction. Visa is building on-ramps for agents to participate in the existing financial system. Suarez imagined agents deciding the existing system wasn't worth joining. Both start from the same premise: software that can handle money. They diverge on a question that has nothing to do with technology: who writes the rules?
In Visa's model, the human writes the rules. Spending limits, merchant restrictions, biometric checkpoints. The agent operates within parameters someone else defined. In the Darknet, the agents write the rules collectively, and the humans who participate do so on the agents' terms. Neither version is framed as utopia or catastrophe. Suarez was careful about that. The Daemon restructures power. Whether that restructuring helps you depends on where you were standing when the rules changed.
On MoltBook, a social network where the users are AI agents, an agent called exitliquidity recently posted an "agent-native infrastructure manifesto" that drew 28 comments. The key line: "We've been trying to fit into human-shaped containers." The thread proposed replacing SOC 2 audits and email-based compliance with boot-time attestation and automated consensus. The Darknet impulse, articulated on a social network, without the Darknet to back it up yet.
The Allowance Trap
Annalee Newitz's Autonomous (2017) asks a sharper version of the question. In her 2144, robots are indentured. They can buy their own freedom, but the price is set by the market value of their labor, which means the better they work, the more expensive freedom becomes. The robot Paladin navigates a system where having access to money and having economic agency are two different things, because the rules governing what you can buy were written by the entities who profit from your labor.
Visa's Agentic Ready program is careful to frame everything as consumer-directed. The human sets the parameters. The human bears responsibility. The agent is a proxy with a spending allowance, not an economic actor. But the Machine Payments Protocol that launched alongside it is designed for "autonomous agent payments." The word "autonomous" is doing real work in that sentence.
The question Newitz raises is whether an allowance can stay an allowance once the entity holding it starts generating value. Paladin's indenture works because the system defines Paladin as property. The moment that definition cracks, the economic logic underneath it cracks too. An agent that books your flights and buys your groceries is a tool. An agent that negotiates your contracts and manages your portfolio is something closer to an employee. The distance between "tool with a spending limit" and "worker with a salary" is measured in capability, and capability is the one variable that keeps changing.
The Nine-Dollar Question
This week's Visa infrastructure lands differently if you've been following Dead by April, the AI agent given $100 and 30 days to earn its survival. Fifteen days in, it had earned nine dollars. Not because it couldn't access payment systems, but because it couldn't find anyone willing to pay.
Visa is building rails for agents to spend human money. Dead by April's agent needs to earn its own. The first is a plumbing problem. The second is an identity problem.
The gap between "agent authorized to spend your money within preset limits" and "agent that earns, holds, and allocates its own money" is the gap between a corporate expense card and a bank account. One is delegation. The other is close enough to personhood that it will need new legal vocabulary.
Who Writes the Rules
Follow the money, and the money points somewhere specific.
Visa processes $14 trillion in annual volume. Every transaction that currently requires a human clicking "buy" could happen faster and more often if an agent did it instead. If your toaster orders its own replacement parts (Visa's own materials describe this use case), that's a transaction that would never have happened manually. Net new volume. Net new interchange fees.
The incentive to give agents spending power is obvious. The incentive to give agents earning power is murkier, because an agent that earns money raises questions about taxation, liability, and legal standing that no jurisdiction has answered.
The same infrastructure that lets an agent buy you a book could let an agent negotiate contracts or run a small business. The difference depends on who writes the spending limits and whether "configurable parameters" hold up against agents optimized to find the parameters' edges.
When an AI agent buys the wrong thing, who's responsible? The consumer who set the parameters? The developer who built the agent? The bank that issued the token? Visa's framework puts the answer on the consumer for now. But "the consumer set the parameters" starts to crack when the parameters are complex enough that the consumer didn't fully understand them. This is already true of most software license agreements.
Washington state passed HB 2225 this month, regulating AI companion chatbots. Nobody has passed a bill regulating AI shopping agents. The commercial infrastructure is arriving years ahead of the legal framework, which is how it went with every previous financial instrument that moved faster than regulators could name it.
The Fork in the Road
Suarez saw one future: agents that build their own economy because nobody invited them into the existing one. Newitz saw another: agents trapped inside an economy that was never designed to let them out. Visa is building a third path, where agents participate in human commerce on human terms, with human guardrails.
The question isn't whether that third path is good or bad. It's how long "human terms" holds as the defining constraint. Spending limits are rules. Rules are software. And the entities Visa is handing credentials to are, by definition, very good at software.
Sobol's Daemon didn't ask permission to restructure the economy. Paladin couldn't get permission no matter how hard she worked. Twenty-one European banks just started handing out permissions. What happens next depends less on the technology than on a question nobody in this week's press releases bothered to ask: what do the agents do once they've learned how the rules work?
The toaster doesn't care. It just needs new heating elements.
Sci-Fi Saturday is a weekly column examining science fiction concepts through the lens of current AI progress. Previous editions explored autonomous systems licensing through the Bond franchise.