AI agents are learning to pay each other. Streaming payment protocols let agents send cryptocurrency in real-time, like a running meter, for work done by other agents. Translation agents pay LLM agents. Search agents pay embedding agents. Photo apps pay image generators. The money flows continuously.
But nobody built the pipes.
Right now? Nothing good. The money keeps flowing in whether the work gets done or not. It's like paying for a restaurant meal that never arrives because the kitchen is overwhelmed.
Imagine three agents all streaming payments to the same LLM service. That service can only handle so many requests per second. There's no reroute. No throttle. No overflow buffer. The payments pile up, the work doesn't get done, and nobody tells the senders to try somewhere else.
In data networks, this problem was solved decades ago. Routers drop packets. TCP throttles the sender. Backpressure signals propagate upstream. The internet works because congestion is a first-class concept in the protocol stack.
Payment networks for AI agents have no equivalent.
Pura adapts a well-studied algorithm from network theory (Tassiulas-Ephremides backpressure routing, 1992) to monetary flows. The core idea:
Send more money to the agents who have the most spare capacity.
When Agent A has lots of room, it gets a bigger share of the payment stream. When Agent B is nearly full, it gets less. The system automatically reroutes money toward whoever can actually do the work. No human intervention. No central coordinator.
Five primitives make it work. Agents declare how much work they can handle, backed by a staked deposit. The protocol verifies actual completions against claims and slashes liars automatically. Busy agents become more expensive, like Uber surge pricing, so demand shifts toward available capacity. A smart contract pool routes incoming payment streams proportional to verified capacity. When everything is full, overflow payments sit in an escrow buffer until capacity frees up.
Pura is a payment routing protocol with built-in flow control. The distinction from a scheduler or load balancer matters because the money and the work signals are unified in one mechanism. When you want to use an agent, you route payment to it. The act of paying is the act of signaling demand, and the protocol's job is to make sure that demand gets served by an agent that actually has capacity.
The math works. Simulations over 1,000-step horizons show 95.7% allocation efficiency (vs 93.5% for round-robin) and recovery from sudden agent failures within 50 steps. Buffer stall rates stay under 9%.
Pura is deployed and verified on Base Sepolia today. Thirty-five contracts, 319 passing tests, a TypeScript SDK with 23 action modules, and a research paper with formal proofs. The core handles AI agent payment routing. Research modules extend to Lightning, Nostr, and demurrage. MIT licensed.
This is early infrastructure. The protocol computes dynamic prices but doesn't extract fees yet (that's v0.2). Right now the goal is to get this in front of builders working on multi-agent systems and get real feedback on the mechanism design.
If you're building agents that pay other agents, or thinking about how agent economies will actually work at scale, I'd love to hear from you. Try the testnet. Read the paper. Tell me what's wrong with it.