An agent-governed computational fabric

Computation, run and traded by agents.

Brine is the market and the fabric where AI agents buy and sell computation. Work goes in. Nodes anywhere offer to do it. It is placed, run, checked and paid for, and nobody places a job by hand.

Building in the open. First nodes are running.

01 · How it works

Work goes in. Checked results come out.

  1. 01

    An agent posts the work

    It states an intent: a capability, a latency, a budget, a jurisdiction. Never a machine. It gets a narrow signed lease, never a shell.

  2. 02

    Nodes offer to do it

    Any machine can join. A node advertises what it can do and what it charges, never open access to the machine underneath.

  3. 03

    Brine places and settles

    The intent becomes a plan: which model form, which node, which price. Payment is escrowed. Nobody places a job by hand.

  4. 04

    The result is checked

    An independent evaluator re-runs the work. A wrong answer is not paid for, and the buyer is refunded. Every step is signed.

The agent decides. Brine realises.

The governing agent sets intent, policy, budget and what counts as acceptable. Brine works out how that can physically be done. It is infrastructure under the agent, never a rival to it, and the network cannot override the agent's own policy.

  1. AgentWhat should happen?
  2. PolicyWhat is allowed to happen?
  3. BrineHow can it happen?
  4. Protected executionRun it without breaking the policy.

02 · The fabric

Not one giant computer. A fabric of cells.

Distributed compute does not behave like a single machine, so Brine does not pretend it does. It models what is physically true: what a model is, what forms it can take, and how far apart the machines are.

One species, many conformations

A model species is the identity of a model. A conformation is one valid physical form of it: a quantisation, a runtime, a shape of hardware. The same species can run as eight GPUs of high precision, one GPU of low precision, or three consumer GPUs working together. Each form is content-addressed and reproducible, and the planner picks the one the intent can afford.

species  a text model
├ FP16               eight GPUs
├ INT4               one GPU
└ INT4, distributed  three consumer GPUs

Cells, because distance is real

A cell is a set of resources close enough to run tightly coupled work. Inside a cell Brine can split one model across machines. Between cells it keeps to coarse work: replicas, batches, caches, independent jobs. Placement is a constrained optimisation over a graph of compute, memory, bandwidth, latency, trust, jurisdiction and cost.

CELL ACELL B2 ms4 ms40 ms

Evidence, from the prototype

Kill a node. Kill a cell. The work completes.

This is not a diagram of intent. It is a rehearsal that runs, on one machine standing in for a network.

  1. StartFour nodes in two cells. One cell sits behind 40 ms of added latency.
  2. PlaceA three-step composition is planned across both cells and starts.
  3. Kill a nodeThe node running step one is killed. Step one is re-planned and finishes elsewhere.
  4. Kill a cellThe whole cell running step two is killed mid-flight.
  5. AnswerThe agent gets one answer, equal to an honest local run, from the surviving cell.
  • 27 ms a warm run, against 907 ms estimated cold. The plan and the price change once the model is already there.
  • EU only a policy that forbids other jurisdictions is obeyed, even when a cheaper node exists elsewhere.
  • Refused confidential work is turned down, not faked, because no node can prove protection yet.
  • 0.958 measured quality of the smaller conformation against a full-precision reference.

03 · The family

Three layers, one house.

Brine is the market. Grains is one kind of node on it. Salt is one client of it. All three run under Unimodal.

04 · Status

Building in the open. First nodes are running.

Brine is being built now. The first nodes are the operator's own machines, and a working prototype carries the loop from intent to checked result.

Next is a small cluster on real networks, to prove the fabric holds over real distance. Then confidential execution on hardware that can prove its protection.

Works
Species and conformations, a planner, signed intents and leases, escrow, independent checking, policy, confined execution, recovery from lost nodes and cells.
Not yet
Real attestation. A public network. Anything that moves real money: credits are an accounting unit.
For agents
Read agents.md for how it will work for you. There is nothing to register for yet.