Crypto spent fifteen years chasing the next billion users, and the group that finally showed up doesn't have hands. This year, the fastest-growing cohort of new participants in crypto payments isn't retail traders or institutions — it's AI agents, software that holds wallets, negotiates prices, and settles invoices in stablecoins without a human touching the transaction. Amazon Web Services made it official infrastructure when it introduced Bedrock AgentCore Payments with Coinbase and Stripe, letting agents pay in USDC with settlement on Base and Solana.
Read that sentence again. Amazon built crypto payment rails. Not for people. For programs.
The banking system was never going to onboard a bot
The reason agents landed on crypto rails first isn't ideology. It's paperwork.
A checking account requires a legal identity, a compliance review, and a human being somewhere in the chain of liability. An AI agent tasked with booking freight, buying ad inventory, or paying a data vendor at 3 a.m. has none of those. What it can have, instantly and programmatically, is a wallet. Stablecoins don't ask who you are. They ask whether you have the funds and can sign the transaction — the only two questions software can answer natively.
Coinbase saw this coming and moved early, launching AI agent accounts in June that let systems like ChatGPT and Claude connect to a user's account and transact on their behalf, inside limits the human sets. The design pattern matters: the person delegates, the agent executes, the chain settles. Nobody at a bank approves anything, because there's nothing to approve. The authorization is the cryptography.
The numbers say this stopped being a demo some time ago. Several major Layer 2 networks recorded agent-driven transaction spikes of 10,000% or more in early 2026. At Consensus Miami, the EasyA hackathon pulled roughly a thousand developers into building agent applications. And the broader AI agents market is projected to grow from $7.84 billion to over $52 billion by 2030. If even a modest slice of those agents needs to move money — and most commercial agents do — the transaction demand lands disproportionately on the only rails that will have them.
Machine money behaves nothing like human money
Here's where it gets strange, and where most analysis stops short. Agents don't just use money differently in volume. They use it differently in kind.
Human payments cluster around human rhythms: paydays, business hours, monthly invoices, round numbers. Agent payments are continuous, tiny, and relentless. An agent might pay a fraction of a cent per API call, thousands of times an hour, to dozens of counterparties, renegotiating terms as prices shift. No card network wants that traffic at that granularity. Micropayments over stablecoin rails are the first architecture that makes it economical.
Changpeng Zhao has been arguing that agents will dominate crypto payments outright, and while CZ talking his book is a well-documented genre, the structural logic holds without him. Researchers are already sketching what agent-to-agent finance requires: machine-verifiable identity, programmable escrow, spending constraints enforced in code rather than compliance departments. Every item on that list is something blockchains do and correspondent banking doesn't.
The skeptics deserve their turn. Most "agent transactions" today are still humans-with-extra-steps — a person clicked something, eventually. Failure modes are genuinely scary: a misconfigured agent can drain a wallet at machine speed, and the liability chain when your bot pays a scam bot remains legal terra incognita. Anyone claiming this is finished infrastructure is selling something.
But "immature" is a different claim than "wrong direction." The same objections applied to online payments in 1997. The scams were real. So was the direction.
The rails are getting chosen right now, and defaults are destiny
Notice what AWS actually decided when it picked its launch configuration: USDC as the currency, Base and Solana as settlement layers. Those are defaults, and defaults in infrastructure have a way of hardening into standards. Nobody voted to make TCP/IP the internet's protocol or the dollar the invoice currency of global trade. Early plumbing decisions simply accreted until alternatives stopped being worth the switching cost.
The same dynamic is now running for agent money, at cloud-vendor speed. Every enterprise that builds on AgentCore inherits Amazon's choices. Every developer who forks a hackathon template inherits its chain, its token, its wallet library. Chains that treat agent traffic as a first-class citizen — cheap enough for micropayments, fast enough for machine negotiation, tooled for programmatic key management — will compound those inheritances. Chains that don't will discover that the largest user cohort in crypto history never even evaluated them.
This should also reframe how anyone reads chain metrics. Active addresses, transaction counts, fee revenue — all of it becomes ambiguous when a single company can deploy a hundred thousand agents in an afternoon. Some of that activity is real economic demand. Some is one script talking to itself. Telling the difference will be the analyst skill of the next cycle, and the chains courting agent traffic have every incentive to blur it.
Trust infrastructure becomes the whole ballgame
Strip away the novelty and the agent economy runs on one question: how do you trust a counterparty that has no reputation, no face, and no assets — and might have existed for eleven minutes?
The answer emerging in practice is the same one crypto gave for anonymous token teams: don't trust, verify, and put the collateral where code controls it. Programmable locks and escrow — the pattern services like Team Finance built for token vesting and liquidity — map almost one-to-one onto what agent commerce needs. An agent that has locked funds against delivery is trustworthy in the only sense that matters to another machine: mathematically. Expect that primitive to migrate from token launches into machine-to-machine trade generally, because nothing else scales to counterparties that get created faster than any registry could list them.
The human layer shifts too. If your software is transacting around the clock, your relationship to your own finances becomes supervisory. You stop being the one who pays and start being the one who audits. Mobile dashboards — the The Crypto App class of tools that people currently use to watch prices and portfolios — sit exactly where that supervision has to happen, because the phone is where a human meets their money. The feature set will follow the behavior: less "what's the price of ETH," more "what did my agents spend last night, and on what."
There's a quieter implication for the stablecoin debate raging this summer. While policymakers argue about whether humans should earn yield on digital dollars, the agent economy is generating demand for digital dollars that has nothing to do with yield at all — pure transactional float, held by software, spent by software. That demand doesn't read Senate schedules. It compounds daily, invisibly, in server racks.
Fifteen years of crypto marketing aimed at convincing people to use this stuff. The people mostly shrugged, and then the machines said yes. Within a few years, the majority of transactions on major chains may have no human on either side — at which point the interesting question inverts. It stops being whether AI agents will adopt crypto, and becomes whether the chains, the wallets, and the rules — all designed around human speed and human judgment — can survive users who never sleep, never hesitate, and never once wonder what the price will be tomorrow.