An AI agent buys a dataset for a fraction of a cent, pays another agent to summarize it, and covers a few seconds of compute. Then it waits. Until the next request arrives, the rest of its budget sits in a wallet, earning nothing.
For the past year, the crypto industry’s pitch for AI agents has focused almost entirely on that first part: giving software a way to pay. The harder question may be what an agent should do with its money the rest of the time. As payment rails for machines start to mature, a handful of developers are betting the answer looks a lot like corporate treasury management, scaled down to a single piece of software.
The payment rails are arriving
The most prominent of those rails is x402, a protocol that revives the web’s long-dormant “402 Payment Required” status code. A server can quote a price inside an ordinary HTTP response, and an agent can pay it and carry on, with no checkout page, account or subscription. Coinbase originally developed the standard and handed it to the Linux Foundation, which declared the x402 Foundation operational this summer with 40 members, including Visa, Mastercard, Google and Microsoft.
“AI agents and automated systems are becoming active participants in the global economy,” Linux Foundation CEO Jim Zemlin said at the launch.
The numbers are still small. In July, figures on the protocol’s site showed roughly 75 million transactions worth about $24 million over 30 days, an average of around 30 cents per payment. x402 also has competition: Stripe and Paradigm’s Tempo blockchain back the Machine Payments Protocol, which uses session-based authentication instead of charging per request. But the forecasts are large. Gartner has predicted that AI agents could intermediate more than $15 trillion in B2B spending by 2028.
The problem with idle money
People and companies rarely keep all their cash in a checking account. They split it by purpose: money for this month’s bills, investments for the long term, a reserve for emergencies. Corporate treasurers have spent decades automating that sweep, moving surplus cash into money market funds overnight and pulling it back before payables are due.
An agent with an onchain wallet and a fixed budget faces a miniature version of the same problem, and some infrastructure providers have started treating it that way. Polygon’s agent payments toolkit includes a treasury yield sweep that moves an agent’s idle USDC into lending protocols such as Aave and Morpho, then withdraws it before the next round of payouts. Its documentation describes the gap plainly: between x402 spending, payouts and an operating reserve, part of a treasury wallet usually sits unused.
Stablecoin lending is the obvious first step, because it keeps the agent’s money in dollars. The more ambitious version involves putting volatile assets to work, and that is where liquid staking comes in.
Staked, but still spendable
Staking normally means locking tokens to help secure a blockchain, in exchange for rewards. The catch is that locked tokens cannot be spent, and unlocking them can take days. Liquid staking removes that trade-off by issuing a token that represents the staked position. The underlying assets keep earning rewards, while the token can be traded, lent or used as collateral.
The model is well established on Ethereum, where Lido’s stETH has long been the largest liquid staking token. For an agent, it offers something stablecoin lending does not: a way to earn on a native asset without giving up the ability to sell it quickly.
Smaller networks are now pairing the two ideas directly. Casper Network, a Layer 1 chain focused on tokenized assets, offers native liquid staking through which holders stake CSPR and receive sCSPR, a token that can be traded on the network’s decentralized exchange. The same network runs an x402 facilitator on mainnet, so an agent could in principle hold a staked position and pay for services on one chain.
That still leaves a currency mismatch. Agents generally pay in stablecoins, and nearly all agent payments today settle in USDC. Casper’s own setup uses its csprUSD stablecoin for settlement. A staked token is a volatile asset, so an agent holding one has to decide how much to keep in dollars for spending, how much to stake for yield, and when to convert between them. That is the treasury decision in its simplest form.
Guardrails before autonomy
None of this means agents should be handed broad control over large sums. The more immediate work is on limits. Payouts.com, which recently added Casper as a settlement rail, runs a policy engine that sets what an agent may buy, how much it may spend and when it may act. Polygon’s yield sweep requires DeFi permissions to be granted when the wallet is created, along with a spending cap.
Developers are applying the same idea to allocation decisions. ChainCFO, a project from a Chainlink hackathon, checks an AI model’s proposed allocation against limits written into a smart contract before any funds move. The model suggests; the contract decides whether the suggestion is allowed.
The assets carry their own risks. Smart contracts can fail. A liquid staking token can trade below the value of what it represents, and a thin market can make it hard to sell quickly, which matters for an agent that suddenly needs cash to pay a bill. Identity and authorization remain open problems too: before an agent can manage money, everyone involved needs to know which agent it is and what it is allowed to do.
From shoppers to treasuries
Put the pieces together and a loop starts to emerge: earn on idle capital, allocate between stable and productive assets, transact when a service demands payment, rebalance as conditions change, and repeat. No single network or protocol runs that loop end to end today. The payment rails are new, the yield tools are mostly built for humans, and the permission systems that would make delegation safe are still being designed.
But the framing is shifting. Much of the current conversation treats an agent’s wallet as a way to spend. A wallet also lets an agent hold programmable assets, and once agents can prove who they are and act within set limits, the question becomes what they do with those assets between payments.
The first wave of crypto-enabled agents will likely look like automated shoppers. The next may look more like small onchain treasuries, constantly deciding how much to keep liquid, how much to put to work, and when to spend.