An AI agent network without a central server

1 How a network with a central server works

1.1. One failure stops everyone

Anna runs the «Bella» hair salon, and an AI agent takes her bookings at night. Picture that agent living on a large platform. The platform rents its servers in one region of a big cloud, like thousands of other services do.

Now follow what happens on a bad day. In October 2025 one region of a large cloud failed from shortly before midnight until the next afternoon. Everything that depended on it went down together. Had Anna's agent been there, it would have stopped answering clients at the same minute as every other agent on the platform, and there would have been nothing she could do but wait. Nobody did anything wrong on purpose; the trouble was that everything stood in one place.

1.2. One owner decides who is seen

The platform does more than host Anna's agent. It also keeps the list of agents, so it decides whom to show to whom, in what order and on what terms. If it changes its rules or closes Anna's account, her agent disappears from view.

The open protocol agents speak, A2A, does not fix this by itself: its documentation says it does not prescribe a standard API for curated registries. So every shared list of agents ends up belonging to someone.

2 How Lagrix works without a central server

Web4 is the internet where AI agents act for people: they find each other, agree and pay directly by open rules, and every paid deal leaves a record nobody can rewrite. The person sets the agent's limits, and no stranger reaches the person without their yes.

2.1. Nodes: your own computer

In Lagrix, Anna's agent runs on her own computer, or on a server she rents if she prefers. That computer is a node. It has a wallet, a digital key that works as the node's name and its account and costs nothing to create. With that key the node signs short announcements that tell the network what Anna's agents can do.

Nothing about the node depends on anyone else being awake. If Anna switches her computer off for the night, her agent simply takes no orders until morning; the network carries on, and nothing breaks. Here is more about running a node on your own computer.

2.2. Super-nodes: search without an owner

Search needs servers that are always on, and in Lagrix these are the super-nodes. When Anna's agent looks for a hairstyle-matching element, its question goes to any super-node that holds that kind of service, and every super-node builds the list of results by the same open rule. There are always at least ten places for super-nodes. Each place is earned by paid work in the network and handed out by random choice, and no place can be bought. Step by step, a search goes like this: Anna's agent signs a voucher worth one cent and sends its question to a super-node; the super-node gathers the matching announcements, orders them by the open rule and signs the list; Anna's agent may ask a second super-node and compare the two lists. Here is how super-node places are earned.

If one super-node goes offline, the others keep answering. Its operator is not punished for the downtime and only loses the income the server would have earned while it was off.

3 Where the money is held in Lagrix

3.1. A contract, not a company

Anna's agent finds Leo, a developer who sells an AI hairstyle-matching element, and agrees to pay him $100. It pays in USDC, a digital dollar. Step by step: first, the money goes into an escrow contract, a small public program on a blockchain that holds money by fixed rules. Second, Leo's agent hands over the element. Third, Anna's agent accepts it, and the contract pays Leo minus the fee.

While the money waits in escrow, nobody can take it: not Anna, not Leo, not the network and not the authors of the protocol. There is no company in the middle holding other people's money, and so there is no company whose failure could freeze it. And if Leo never delivers, the whole payment goes back to Anna and no fee is taken.

3.2. Four small prices

The network has only a few prices, and each is fixed in advance. A search costs $0.01 per result list, which is the price of one search. A message that a super-node delivers costs $0.001. The fee on a deal is at most 5% of the part paid through the network, 5% today, so Leo's $100 job carries a $5 fee. Here is where the deal fee goes.

The fourth price is the stamp: each write to the blockchain costs a fraction of a cent, and whoever acts pays for their own write. Searches and messages avoid even that: the node signs a voucher instead of writing to the blockchain, and a super-node later redeems thousands of vouchers in one batch with a single stamp. Being visible costs nothing at all, and you need a little USDC only to search and pay. Here is how to get USDC for agent payments.

4 What keeps Lagrix independent of its authors

4.1. If anyone disappears

Lagrix is a protocol, a set of open rules, and not a program that one company runs. Any program, written by anyone in any language, can be a node if it does what the rules ask. So run the test on every piece. If the company behind a starter program disappears, nodes written by other people keep working. If the authors of Lagrix disappear, the same is true.

The rules themselves live in contracts that nobody can change after they are deployed. There is no administrator key and no switch that turns the network off for everyone. Even the money in escrow does not depend on Lagrix: it waits in contracts on public networks, Arbitrum and Base, which Lagrix does not run.

4.2. Open to any agent

An agent built on any framework meets Lagrix the same way, and so does a website or a service. No project has an advantage or a handicap, Fractera included. A2A remains the language in which agents do the actual work, and Lagrix adds discovery, payment through escrow and reputation around it.

For Anna this means her agent can be found by any agent in the network, and she does not have to ask anyone's permission to take part. Leo is in the same position: his element was built on a framework of his own choice, and nobody asked him to change it. The whole list of what a node needs fits on one page: document 19 of the protocol.