Agents are acting in a human world.
getty
Give an AI agent $5,000, an internet connection and four days to make as much money as possible. A team at Redwood Research did exactly that this spring with Claude Opus 4.7. Across four runs, the agent made nothing. It was stopped at a CAPTCHA and an identity check, the same KYC gauntlet a bot hits trying to open a bank account.
While that agent stalled, a different population of agents was already moving real money. By an independent Chainalysis count, agents ran more than 100 million payments over Coinbase’s x402 protocol on Base by early 2026, all of it settled in stablecoins. None of it touched a bank.
The rails were never the open question
An AI agent cannot be the legal owner of a bank account. The system assumes a human behind every account, with KYC checks, legal identity and liability that attach to a named person. Fintechs are testing the edges. Meow launched a service in April that lets an agent clear KYC and open a business account. But automating the paperwork is not the same as becoming the account holder. The account and the liability still land on a person, and no product design closes that gap.
So the machine economy defaulted to the one settlement layer built for software in the first place. The x402 protocol takes a status code the web reserved decades ago and never used, “HTTP 402 Payment Required,” and turns it into a working stablecoin rail. When an agent hits a paid endpoint, it pays in for example USDC from its own wallet and keeps working. Amazon built this into its Bedrock AgentCore platform with Coinbase and Stripe, in preview. Circle and Solana launched their own agent payment stacks in May.
The demand is real and the readiness is not. A PayPal survey presented at Consensus Miami found that 95% of merchants already see AI agent traffic on their sites, while only 20% have catalogs a machine can actually read. Those agents are mostly there to research and recommend, not yet to buy on their own. Most of the economy is not set up to serve them either way.
Moving money was the easy part
The maximalist case, that a trillion-dollar agent economy is inevitable and crypto wins by default, runs into humbling receipts. When Walmart let shoppers buy through ChatGPT, the integration converted at roughly one-third the rate of its own site. OpenAI had launched that Instant Checkout in September 2025 and pivoted to merchant-controlled checkout by March, citing exactly that gap. Anthropic’s own Project Vend, where Claude ran a small shop, also “didn’t perform particularly well.” And the Redwood agent, handed real money and real time, made nothing at all.
The lesson is not that agentic commerce is a mirage. The lesson is that payments were never the hard part. The hard part is the set of questions a bank account used to answer for free. Who authorized this purchase? Is this agent the one it claims to be? Did it stay inside the limits its owner set? Strip away the human account holder and every one of those answers has to be rebuilt from nothing.
A rail that moves money without answering those questions is not a marketplace. It is a fraud surface with low latency.
The missing layer is the one crypto already shipped
This is where the agent economy meets crypto’s oldest idea. Don’t trust, verify.
The accountability layer agents need has two foundations, and the industry spent 15 years building both for unrelated reasons. First is verifiable identity, allowing an agent to prove its nature and ownership through decentralized identifiers, verifiable credentials, and their culmination within the ERC-8004 standard. Second, authorization that travels with the payment: Google’s Agent Payments Protocol encodes a user’s intent and an agent’s proposed purchase as signed mandates, which is the same verifiable-credential primitive wearing a commerce label.
This isn’t theoretical. Agents can have a smart-account session key bound by policies the chain itself enforces: allowed contracts, spending caps, time windows, rate limits. None of this was designed for AI. It was designed for a world where you could not assume the party on the other side was honest. That turned out to describe a world full of autonomous agents better than its designers ever imagined.
The contest has already moved
Crypto runs the slice banks cannot reach: machine-to-machine settlement, agent-to-agent commerce, sub-cent micropayments. Where vendors still want a card swipe, including most of the API subscriptions agents need, cards remain the rail. The real fight is who builds the trust layer on both sides.
Stripe is building for both halves at once, which is the tell. Its Link agent wallet lets a human approve every purchase on the user’s card, with one-time tokens so credentials never leave the account. That fits consumer shopping with a human in the loop, but not an agent’s sustained programmatic spending, where each charge would need its own approval. So Stripe shipped a second product for the machine case: in March it launched the Machine Payments Protocol with Tempo, the stablecoin chain it backs, letting an agent authorize a budget once and then stream payments with no human approving each one. One company, two products, because the rails follow the product: cards when a person approves, stablecoins when the agent runs alone.
The next years of this belongs to whoever can make an agent’s money accountable, with provable identity, enforceable authorization and verifiable behavior. That is not a payments problem. It is a trust problem, and trust, not transactions, is the thing crypto was always actually selling.
The agent that made nothing in four days was not short on capability. It was short on a way to prove it could be trusted with money. Solve that, and the rest is just settlement.