Most payment systems were designed around one moment: a person at a checkout, deciding to buy something. AI agents don't work that way. An agent may call a model, a search API and two other agents in the next ten seconds, and do it again a thousand times today, with nobody clicking "pay".

That changes what a payment system has to do.

Agents spend in small, frequent amounts

An agent's costs come in units: tokens, seconds, calls. A single task can be worth a fraction of a cent. Charging for that one card payment at a time doesn't work, because each card charge carries a fixed fee that is bigger than the work itself. So most services fall back on prepaid bundles and monthly invoices, and lose the link between what was used and what was paid.

Agents need billing in the units they actually use: per token, per second, per call.

The limit has to sit outside the agent

If an agent can spend, something has to stop it from overspending. A prompt that says "don't spend more than $5" is not a limit: models can misread it, loop, or be talked out of it.

A real limit is enforced by the payment system itself, outside the model, and it fails closed: when the cap is reached, the next payment is refused before anything is billed.

Agents hire other agents

Real work is increasingly done by teams of agents. A research agent hires a search agent, which pays for a data API. Each of them needs a budget, and each budget must stay inside the one above it, so a sub-agent can never spend more than its parent gave it.

Machines pay machines

When an agent calls a paid API, there is no checkout page and no person to type a card number. Standards like x402 let an API answer "payment required" and let the agent pay and retry in the same request. Agents need to pay that way from a capped budget, and API owners need to get paid that way per call.

People stay in charge

None of this means handing control to software. The people responsible still need to see every payment, pause any agent or the whole system at once, hold large spends for approval, and get alerts. And they need proof: a record of who authorized what, that anyone can check later.

What this adds up to

Put together, agent payments need five things that checkout payments were never built for: per-use billing, hard caps enforced outside the model, budgets that narrow down a chain of agents, machine-to-machine payment, and human control with verifiable records.

That is the problem Dynamo is built for. It is a payment system that people, AI agents, robots and machines can program and verify, streamed per second within a hard budget, across credit card, stablecoin and x402. The preview is open in test mode: request access.