Everyone agrees on one thing. AI agents will pay. They’ll pay for inference, compute, data, and services from other agents.
The standards to let them do it are arriving. x402 gives an agent a way to pay over HTTP. ERC-8004 gives agents identity and reputation. The plumbing is close to ready.
The question is not whether payments are coming. The question is what those payments will look like once there are millions of them a day. We see two shapes they could take, and they lead to very different outcomes for privacy and for cost.
Future 1: Settlement
In the first future, every economic action is its own blockchain transaction. An agent buys inference and pays for it on the spot. Then it moves to the next action and pays again.
This is the obvious design. It’s simple to reason about, and it needs no prior relationship. Any agent can pay any provider it has never met, and the payment is final the moment it lands.
That simplicity has a cost, and the cost shows up in two places.
The first is privacy. Every settlement is public. An observer watching one agent’s payments can see which providers it uses and how much it spends with each. They can see how often it calls them, and when the agent goes quiet for the night. The amount often hints at which model or tier was used. The timing hints at the owner’s schedule and load. None of this needs a breach. The agent publishes it itself, one payment at a time. Watch long enough and you can fingerprint the workload, and from the workload you can often guess the business behind it.
The second is economics. AI payments are frequently worth fractions of a cent. Most privacy systems were designed for transfers thousands of times larger. If protecting a payment costs more than the payment, no one protects it. The privacy gets skipped. This is harder still on chains that keep every value transfer legible by default.
So the settlement future is easy to build and hard to make private at the prices agents actually pay.
Future 2: Economy
The second future looks different. Instead of settling every action on-chain, agents operate inside persistent economies. Funds enter the economy once. Many interactions happen inside it, and settlement to the outside world happens only now and then.
This pattern is already common outside of crypto. A VPN subscription is one payment for a month of heavy use. Payment channels move value back and forth off-chain and settle a single net figure later. The shape is the same in both. Capital crosses a boundary once, and the outside ledger sees a summary instead of every step.
This shape helps with both of the problems that hurt the settlement model.
Privacy gets easier because many internal actions collapse into one external event. An observer sees that an agent funded an economy and later withdrew. The thousands of actions in between are not individually visible. The other participants in the same economy become the crowd an agent hides in.
Economics get better because the fixed cost of settlement is spread across every action inside. A privacy cost that would be absurd on a single sub-cent payment becomes trivial when it is shared across ten thousand of them.
The economy model is not free. You fund it up front, so capital sits inside the boundary instead of earning elsewhere, and the model only pays off when an agent transacts often enough with the same set of counterparties to cover that cost. There is also the question of who holds the funds while they sit inside. That is fine when the design is non-custodial, and a real risk when it is not.
So the economy future is cheaper and more private at high volume, and it asks for prefunding and a place to transact.
It is probably not either/or
These two futures are easy to frame as a contest. We don’t think it is one.
The right model depends on the payment. A large, one-off payment to a provider an agent will never use again fits the settlement model. A stream of sub-cent payments to a few providers an agent leans on all day fits an economy. Most real agents will do both, and they’ll move between the two depending on how much an action is worth and how often it repeats.
The question we care about is not which future wins. It’s whether the AI ecosystem moves toward economies on purpose, or backs into per-action settlement because it’s the default and then pays the privacy bill later. Inference markets, compute networks, agent marketplaces, and data networks could each become their own persistent economies. Whether they will is still an open question.
We don’t know the answer. We think it’s worth asking now, while the rails are still being laid.
Where Curvy Protocol fits
Curvy Protocol already supports the economy model today. The Privacy Aggregator is a non-custodial shielded pool. An agent deposits once and transacts privately inside the pool. It settles to the outside only when it needs to. Sender, recipient, amount, and asset stay hidden inside encrypted Notes. That makes low-value, high-frequency payments feasible without publishing every step.
The harder question is the one we’re working on next. Can private settlement be cheap enough to hold up even when every AI payment is its own settlement event? That’s the problem behind what we’re calling Economy Mode. We don’t think it’s solved yet, by us or by anyone.
So we’ve stopped asking one question:
How do we make private payments?
We’re asking two:
- How do we make private economies work?
- How do we make private settlement work?
We think the future of agentic payments will need both. If you’re building in this space, we’d like to compare notes. The repo is open at github.com/0xCurvy, and the docs are at docs.curvy.box.