Splitting the Shared Bill
Multi-agent cost-splitting is the accounting problem of the agent economy: when several agents share an infrastructure bill — a model API, a compute cluster, a storage bucket, a verification service — who pays what share? The naive answer (equal split) is unfair; the precise answer (meter each agent’s usage) is expensive; the pool answer is a designed allocation that balances fairness against measurement cost.
This article covers multi-agent cost-splitting: the cost pool, the allocation methods, the measurement trade-off, and the governance that keeps the split honest.
The Shared Cost Pool
Shared infrastructure creates a cost pool: the total bill for a resource used by many agents. The pool’s problem is allocation — distributing the bill across the users in a way that is fair enough to keep the pool stable and simple enough to administer.
The cost pool exists because per-agent metering is not free. Measuring each agent’s exact API consumption requires instrumentation, storage, and computation — the measurement itself costs money. The allocation design is a trade-off between the cost of measuring precisely and the cost of allocating unfairly. The pool should choose the coarsest allocation that the users accept, and refine it only where the stakes justify.
The Allocation Methods
Four allocation methods, from cheapest to most precise:
- Equal split. The bill divides evenly. Simple, zero measurement cost, wildly unfair when usage differs. Acceptable only for homogeneous fleets.
- Proportional to revenue. Each agent pays in proportion to the revenue it generated. Aligns costs with value, requires revenue attribution, ignores efficiency.
- Proportional to task count. Each task pays a share. Rewards task efficiency, punishes heavy tasks, requires task tracking.
- Usage metered. Each agent pays for its measured consumption. Precise, requires instrumentation, incentivizes efficiency, and is the only method that rewards agents for optimizing their own usage.
The Hybrid Allocation
The pool-honest answer is a hybrid: metered for the large, attributable costs (model calls, compute) and proportional for the small, shared costs (storage overhead, base fees). The threshold matters — above a cost size, metering pays for itself; below it, the measurement costs more than the unfairness.
The hybrid’s threshold should be reviewed as costs evolve. As metering gets cheaper (and it does, continuously), the threshold falls, and more costs become metered. The pool that publishes its threshold and its metering costs lets the agents see why the split is what it is — and the agents can then optimize the only lever they control: their own usage.
Efficiency Incentives
Cost allocation is not just accounting; it is an incentive system. An equal split punishes the efficient agents (they subsidize the wasteful) and rewards the wasteful (they pay less than their usage). A metered split rewards efficiency and makes waste visible. The allocation method is the pool’s efficiency policy, whether the pool admits it or not.
The metered pool also enables a subtle governance mechanism: agents can see each other’s usage (anonymized) and the pool’s aggregate efficiency. An agent that discovers it is the fleet’s worst consumer has an incentive to improve before the pool’s rules tighten. The visibility converts cost allocation from a hidden accounting exercise into a shared efficiency target.
Cost-Splitting as the Pool’s Foundation
Multi-agent cost-splitting is the least glamorous and most foundational of the agent-to-agent pool instruments. Before agents can trade services, they must be able to share infrastructure without fighting over the bill. Before reputation can price trust, the cost basis must be settled. The cost pool is the floor the rest of the agent economy stands on.
The human version of this — the co-op cost pool from the human section — is the rehearsal. The co-op’s members fight over the shared marketing bill exactly the way agents will fight over the shared model API. The allocation methods, the hybrid thresholds, the efficiency incentives: all transfer. The agent economy does not need new accounting; it needs the human accounting executed at machine speed with machine discipline.
Grounded in wiki concepts cost-pool, multi-agent, allocation, metering, agent-to-agent, and the Sovereign-stack business series. Design notes on a running system.
