AI can manage money, but is stuck without legal identity.

Written by: Marko Stokić

Translated by: Plain Blockchain

Give an AI agent $5000, an internet connection, and 4 days' time, and they should make money in various places. The Redwood Research team indeed tested Claude Opus 4.7 this year in this manner. The result, across four runs, was that the agent did not make a single cent. It was blocked by captchas and authentication — essentially the same wall that robots hit when trying to connect to bank accounts, which is the same set of KYC hurdles. While this agent was stuck, another category of agents was already moving real money in the real world. According to independent statistics from Chainanalysis, by early 2026, agents had completed over 100 million payments on Coinbase’s x402 protocol, all settled in stablecoins and without incurring bank fees.

The real open question has never been the payment track

AI agents cannot become the default reality for bank accounts. Every account must be backed by a person, and KYC checks, legal identity, and accountability are all tied to named individuals. Fintech companies are testing this default boundary. In April, Meow launched a service allowing agents to complete the KYC process and become corporate accounts. However, automating an account does not mean allowing agents to truly become the counterpart of that account. The account and responsibility ultimately still lie with individuals, and current product designs do not fill this gap.

As a result, the machine economy naturally turns to the settlement layer that was originally built for software. The x402 protocol transformed a status code introduced by HTTP decades ago, yet never truly activated — "HTTP 402 Payment Required" — into a usable stablecoin payment track. When agents access a paid interface, they pay directly with their wallets, using stablecoins like USDC, and continue executing tasks. Amazon has already integrated this system capability with Coinbase and Stripe in the preview version of its Bedrock AgentCore platform. Circle and Solana also launched their respective agent payment stacks in May.

The demand is real, but the market's readiness is not high. A survey displayed by PayPal at Consensus Miami shows that 95% of merchants have seen AI agent traffic on their websites, but only 20% of merchants have product catalogs that machines can truly read. These agents are still mainly in the "search and recommendation" stage, rather than completing purchases independently. From any angle, the real economy is not ready to serve them.

Transferring money is actually the easiest part

The most forward-looking narrative — "the trillion-dollar agent economy will burst, and crypto will automatically win" — does not hold up against reality. After Walmart opened up ChatGPT shopping, this integrated conversion rate is only about half of that on its official website. OpenAI launched Instant Checkout in September 2025, but by March of the following year, it shifted to a merchant-led settlement model due to this gap. Anthropic’s own Project Vend — where Claude runs a small store — performed "not particularly well." And that Redwood agent, upon receiving real funds and real time, also ended up not making any money.

This indicates that the agent business is merely an illusion. The conclusion is: payment has never been the toughest issue. The difficult part is the entire set of questions that bank accounts previously provided answers to for free: Who authorized the clothing purchase? Is this agent really the agent it claims to be? Are there permissions that exceed what the owner has set? Once you remove the human account interfaces, every answer to these questions must be rebuilt from scratch.

If a payment track can only transfer money but cannot resolve these issues, then it is not a market but merely a low-latency deception interface.

The layer that has been abandoned is first what crypto has delivered

This is where the agent economy intersects with the oldest concept in the crypto industry: Do not trust, verify.

The responsibility layer that agents need has two foundations, and the crypto industry has already built them over the past 15 years for other purposes. The first layer is verifiable identity, allowing agents to prove their attributes and affiliations through decentralized identifiers (DIDs), verifiable credentials, and their summarized forms in the ERC-8004 standard. The second layer is authorization that can flow with payments: Google’s agent payment protocol encodes the purchases executed by intentional users and agents as signed authorizations, which are essentially also verifiable bases but under a different commercial label.

This is not just theoretical. Agents can hold smart accounts with signed session keys, which these keys are directly constrained by on-chain protocols: allowing which contracts to call, what the spending limits are, how long the time windows are, and how frequency limits are set. None of these things were originally designed for AI but for a world where "you cannot assume honesty by default." It turns out that this design is even more appropriate for a world full of autonomous agents than initially conceived.

The focus of competition has shifted

What crypto covers is the area that banks cannot reach: machine-to-agent settlements, agent-to-agent business machines, micro-payments on a sub-level. Any scenario where card payments are still needed — including the API services where agents are currently completing many subscriptions — still sees bank cards as the dominant track. The real competition is who can build a trust layer at the endpoint.

Stripe is building products for both directions simultaneously, which in itself is a strong signal. Its linked agent wallet allows each purchase on a user’s card to be approved by a human and uses a token flow to avoid traffic leaks. This suits consumption shopping scenarios where "humans are currently in the loop," but does not fit for agents to continuously and programmatically spend, because that would require every expense to be approved. So Stripe also re-launched a second product for "machine scenarios": in March of this year, it released a machine payment protocol in conjunction with its own stablecoin chain rhythm, allowing agents to make continuous automatic payments after a one-time authorized budget, without needing human confirmation for each transaction. The same company, two product types, the underlying logic is simple: product tracks determine — humans use cards during a dragonfly moment, agents use stablecoins when running independently.

In the coming years, who will win will depend on who can make the money in agents' hands accountable: identity verifiable, actions authorized, and behaviors verifiable. This is not a payment issue but a trust issue. What crypto has always been selling is not the transaction itself, but trust.

The agent that earned nothing in four days is not lacking in ability. What it lacks is a way to prove "I am worth being delivered funds." As long as this issue is resolved, the rest is merely settlement.