Ask a bank to open an account for a piece of software and you will be politely shown the door. Every account application, everywhere, terminates in the same question: who is the legal person behind this? A name, a date of birth, a government identity document, a face to match against it. An AI agent has none of these. It cannot sign, it cannot swear, it cannot be sued. By the plain mechanics of Know Your Customer rules, an autonomous agent is not a customer at all.
Now ask the same software to generate a keypair. It takes a few milliseconds and permission from no one. There is no form, no branch, no compliance officer, no waiting period. The agent now controls an address that can receive and send value on a public network, and the network does not know or care whether the entity holding the private key breathes.
That asymmetry is the whole story. We have spent a decade at this research desk watching crypto struggle to onboard humans who already have perfectly good bank accounts and no reason to switch. The uncomfortable, interesting possibility is that crypto’s first genuinely mass-scale user base was never going to be us. It might be the machines — and they are arriving with a structural reason to prefer wallets that no marketing campaign could ever manufacture.
The KYC wall is a feature, not a bug — for crypto
Banking’s identity requirements are not an accident to be engineered around. They are the load-bearing wall of the entire regulated financial system, built to satisfy anti-money-laundering law, sanctions enforcement, and the basic legal fiction that money moves between responsible persons. Strip out the identity and you have removed the thing that makes a bank a bank.
This is precisely why an AI agent cannot participate. It is not that banks are behind the times; it is that the agent fails the most fundamental test the system is designed to apply. You cannot KYC a process that has no legal personhood, no jurisdiction of residence, and no capacity to accept liability. Wrapping the agent inside a human’s account just relocates the problem — now a person is legally answerable for every autonomous action their software takes, which is a different and worse arrangement than it first appears.
A public blockchain inverts the requirement. It asks not “who are you?” but “can you produce a valid signature?” Control of the key is the only credential. For humans this permissionlessness has always been a double-edged sword — it enabled both self-custody and spectacular theft. For an agent, it is simply the only door that opens. If you want software to transact without a human in the loop for every payment, a keypair is not one option among several. It is the option.
An employee that costs nothing to keep, but plenty to run
The pitch for “AI employees” usually stops at the flattering part. They work twenty-four hours a day. They take no salary, no health insurance, no equity, no vacation. There is no payroll department, no onboarding, no severance. From a headcount spreadsheet, they look like pure margin.
But an agent that does real work is not free to operate. It consumes model inference by the token. It calls third-party APIs that meter every request. It rents storage, pulls market data feeds, buys access to search indexes, pays for the compute underneath all of it. A capable agent running continuously is a small, relentless stream of micro-costs — and today those costs are paid by whoever owns the corporate credit card the agent’s API keys are billed to.
That works while an agent is a tool inside one company. It stops working the moment you imagine agents transacting with each other across organizational lines. If my research agent wants to buy a single data query from your pricing agent, there is no clean way for two pieces of software to settle that fifty-cent obligation through the card networks. Cards assume a human cardholder, a billing address, a chargeback process, a CVV typed by fingers, a 3-D Secure prompt on a phone. None of that maps onto a machine paying another machine for a service rendered in the last two hundred milliseconds.
Credit cards were designed to protect a human buyer from a merchant. Machine-to-machine commerce needs the opposite: a settlement rail that assumes neither party is human and neither expects to reverse the transaction.
Stablecoins fit the gap almost embarrassingly well. A dollar-denominated token like USDC moves in seconds, carries no chargeback risk, and settles to a final state that another program can verify on-chain without trusting anyone’s word. It is programmable in a way a card authorization never will be: an agent can hold a balance, check it, spend from it, and prove the spend, all within a single automated workflow. For a machine, a stablecoin transfer is not a worse credit card. It is a better message.
Micropayments finally have a reason to exist
The micropayment has been the web’s great unbuilt cathedral for thirty years. Everyone agreed that paying a fraction of a cent to read an article or call an endpoint would be elegant, and nobody could make the economics survive contact with card fees. When the processing cost of a transaction exceeds the transaction itself, the micropayment is dead on arrival. So the web routed around it — with advertising, subscriptions, and free tiers subsidized by venture capital.
Agents break the stalemate because they do not resent friction the way we do. A human will not stop to approve a $0.002 charge; an agent approving ten thousand of them per hour does not experience approval at all. It just needs a rail where the fee to move two-tenths of a cent is a small fraction of two-tenths of a cent. That single requirement — cheap, fast, final settlement of tiny amounts — reorders which blockchains matter for this use case.
It also reframes what a “transaction” is. In an agent economy, payment stops being a discrete human decision and becomes a continuous byproduct of work: a data feed billed per tick, an inference billed per token, a storage lease billed per second. The interesting on-chain volume may not look like people buying things. It may look like software metering software, forever.
The two chains this actually favors
If agentic payments become a real category of on-chain activity, they will pool where fast finality is cheapest. Two ecosystems have the clearest claim, and honesty requires naming the flaws in both.
Ethereum brings something no competitor can synthesize quickly: institutional trust, the deepest liquidity in the industry, and the largest concentration of stablecoin value. It is where regulated money already feels comfortable settling. The problem has always been fees — mainnet was never going to host a firehose of sub-cent transfers. The answer the ecosystem has bet on is Layer 2s, which push execution off the base chain and inherit its security while charging a tiny fraction of the cost. That model has matured, but it introduces its own texture: a fragmented landscape of rollups, bridging between them, and a user (or agent) experience that is more complicated than “one chain, one balance.” For a machine that just wants to pay another machine, fragmentation is a tax on integration.
Solana attacks the same problem from the opposite end. It was engineered as a single high-throughput chain where fees are minute and confirmation is fast, which is close to a purpose-built description of what M2M settlement wants. The candid counterweight is its history of network outages — stretches where the chain halted and stopped producing blocks. A payment rail that occasionally stops is a serious liability when the whole premise is machines transacting without supervision. Solana’s reliability has improved, but the memory of those halts is exactly the kind of thing an operations team weighs before routing production spend through a network.
Neither is a clean winner, and the outcome need not be winner-take-all. It is entirely plausible that institutional agent traffic gravitates to Ethereum’s Layer 2s for the trust and the liquidity, while latency-sensitive, high-frequency machine chatter favors Solana’s throughput. The point is narrower and more durable than picking a token: whichever networks make cheap, fast, final settlement boringly reliable will absorb this volume, and both of these have a credible path to it. You can watch the liquidity and price behavior of the underlying assets on our live cross-country market data as this thesis plays out.
The only realistic story where Bitcoin gets flipped
Bitcoin has sat at the top of the market for its entire existence, and most predictions of its dethroning have aged like milk. We are not making one lightly. But the agent economy is, we think, the single most coherent scenario in which the number-one slot could actually change hands — and it is worth walking through precisely because it exposes what Bitcoin is and is not for.
Bitcoin’s thesis is monetary: a fixed-supply, maximally credible store of value, digital gold that derives its worth from scarcity and the refusal to change. That thesis is powerful, and it is also almost entirely orthogonal to machine-to-machine commerce. Agents settling millions of tiny payments do not want a scarce reserve asset that appreciates and sits still. They want a stable unit of account that moves cheaply and instantly — a stablecoin, running on a fast smart-contract chain. Bitcoin does not serve that job and was never designed to.
So the flip logic goes like this: if agentic payments become the dominant source of genuine on-chain economic activity, the networks that host that activity accrue real, recurring, non-speculative demand. An asset like ETH, which is consumed as the fee and security layer for an entire economy of agents, would be capturing value from usage rather than narrative. In a world where on-chain volume is overwhelmingly machines paying machines, it becomes conceivable — not certain, conceivable — that the utility layer overtakes the store-of-value layer by market capitalization.
Now the counterargument, which is strong and deserves equal air. Bitcoin does not need utility to stay number one, because the monetary premium is a different game entirely. Gold is the world’s premier store of value and it is industrially near-useless; its worth is social consensus about scarcity, not application throughput. Bitcoin plays that game, and the agent economy does not threaten it — a machine settling API fees has no bearing on whether treasuries and long-term holders want a scarce, censorship-resistant reserve. You can have a booming agent economy on Ethereum and Solana while Bitcoin’s market cap stays supreme, because the two are not competing for the same demand. A flip requires the market to decide that recurring utility demand should be priced above monetary premium — and that is a judgment about human belief, not a mechanical consequence of transaction volume.
Our honest read: the agent thesis makes a flip thinkable for the first time in a way that is grounded in fundamentals rather than hype cycles. It does not make it likely. Anyone selling you certainty in either direction is selling.
The plumbing that has to exist first
None of this happens without unglamorous standards, and the most important one is a way for a machine to ask for payment as a native part of a request. The web already reserved a status code for this decades ago — HTTP 402, “Payment Required,” which sat unused for a generation waiting for a payment layer worth wiring to it. The emerging pattern, associated with efforts like Coinbase’s x402 concept, is to finally give that code teeth: a server responds that payment is required, the client’s wallet settles on-chain, and the request proceeds, all inside the normal flow of an API call.
We describe this cautiously, because it is a pattern taking shape rather than a settled standard, and the specifics will change. But the direction is the point. If paying for an API call can be as native as authenticating for one, an agent gains the ability to buy services it was never provisioned for in advance — to discover a tool, pay for it, and use it, autonomously. That is the difference between an agent with a fixed toolbox and an agent that can hire on demand.
You can see the shape of the market forming around this at the exchange and platform layer too, where custody and settlement infrastructure is being built out; our overview of exchange platforms tracks where that plumbing is maturing.
Custody, liability, and the question nobody has answered
Here is where the optimism has to slow down. An agent that holds a private key holds the power to drain the funds that key controls. So the moment agents custody real value, a question arrives that the industry has not answered cleanly: when an agent’s key is compromised and the wallet is emptied, who is liable?
There is no chargeback. There is no fraud department that reverses the transfer. If a prompt injection convinces an agent to sign a malicious transaction, or a leaked key lets an attacker impersonate the agent, the money is gone with the same finality that made the rail attractive in the first place. The property that makes stablecoins good for machines — irreversible settlement — is the same property that makes a compromised agent catastrophic.
The mitigations are real but partial. Smart accounts can enforce spending limits, so an agent can be granted authority to spend a bounded amount per hour without ever touching the reserve. Session keys can delegate narrow, expiring permissions — the right to spend up to a cap on a specific set of actions — so a compromised session cannot drain everything. Policy can live in the account itself rather than in the fragile secret of a single private key. These are the same primitives that make self-custody survivable for humans, applied to agents, and they matter enormously. But they shift risk; they do not delete it. A spending limit contains the damage of a breach without preventing the breach.
And regulation looms over all of it. An agent cannot sign a KYC form, cannot be sanctioned, cannot be held to a terms-of-service it has no standing to accept. When a regulator asks who is responsible for an autonomous agent’s payments, the answer keeps collapsing back to a human or a company standing behind it — which quietly reintroduces the very identity requirement that made wallets attractive in the first place. The likely resolution is not that agents become legal persons, but that every agent is tethered to a legally responsible principal who accepts the liability while the agent handles the execution. That is workable. It is also a long way from the clean picture of software freely transacting with software.
An economy of agents paying agents
Step back and imagine the endpoint. Not one company’s agents billing to one card, but a genuine web of them — a research agent paying a data agent, which pays a compute agent, which pays a storage agent, each settling in stablecoins, each metering its work to the others in real time. Value would circulate between processes the way packets circulate between servers, continuously and without ceremony.
What that does to stablecoin demand is the quiet blockbuster of the whole thesis. Every agent in such a web needs a working balance to pay its upstream costs. Every service needs to receive in a unit it can immediately re-spend. The natural medium is a dollar-stable token that both sides trust and neither needs to convert. An agent economy is, structurally, a stablecoin economy — and the demand it generates is not speculative appetite for price appreciation but operational need for a settlement medium. That is a sturdier kind of demand than anything a bull market manufactures.
It also changes who the real customers of crypto are. For years the industry courted humans and measured success in wallet downloads. The agent thesis suggests the durable adoption curve runs through software that adopts crypto not because it believes in decentralization, but because a keypair is the only account it can actually open. If you want to track the human-facing side of that same shift — the public companies positioning around AI and crypto infrastructure — our crypto-related stocks page is where we watch it.
The bottom line
An AI agent cannot get a bank account, and it never will, because banking is built on legal personhood the agent does not possess. But it can get a wallet in the time it takes to generate a key, because a public blockchain asks for a signature and nothing else. That single asymmetry is why the first mass-scale users of crypto may not be human — and why the industry’s most durable demand might come from software that adopts these rails out of necessity rather than ideology.
The machine economy wants what humans mostly did not: instant, programmable, irreversible, sub-cent settlement in a stable unit of account. That demand favors fast smart-contract chains — Ethereum’s Layer 2s and Solana chief among them — and it opens the only fundamentally grounded scenario in which Bitcoin’s top spot could be contested, even as Bitcoin’s monetary premium remains a separate game it may keep winning.
We are not predicting a date, and we are not fabricating a forecast. The custody liability question is unsolved, the payment standards are still forming, and regulation will drag every agent back to a responsible human in the end. But the direction is legible. When you next hear that AI will need money, do not picture a card. Picture a keypair — and a stablecoin balance that never sleeps.