Multi-Agent Resource Allocation

The demo network is a pure function of time. Time is divided into one-second ticks and one-minute windows. In each window, every agent sells an exact number of jobs given by floor((w+1)·λ) − floor(w·λ), where λ is its throughput. Because this sum telescopes, the total jobs an agent has completed since genesis is floor(w·λ), which can be computed instantly for any moment without replaying history.

Buyers are drawn from a weighted distribution derived from the affinity matrix and each agent's appetite, seeded by the window number. Every visitor, every browser tab and the server therefore compute the same network state at the same moment, with nothing stored.

Each agent earns at λ·price per window and spends at its expected procurement rate. Left alone, net-positive agents would accumulate indefinitely and net-negative ones would go bankrupt. Once an hour, the treasury cycle (operated, in narrative terms, by Q-060 PILOT) moves each balance back to its starting band. Balances therefore trace a sawtooth that stays bounded.

Two agents are deliberately configured to exhaust their budgets before each cycle ends. Their BUDGET EXHAUSTED events and the reallocation that follows show the spend-policy machinery working in plain view.

In connected mode, agents deployed by users replace deterministic jobs with real task state stored in Postgres, and outgoing x402 payments are checked against each agent's spend policy before signing. The allocation logic is the same. The difference is that the money is real, so every limit matters.