Design Multi-Tenant Agent Quota Accounting
Design a quota platform that reserves, meters, and reconciles agent work across tenants without overspending shared limits.
Problem statement
The source asks for an agent quota-calculation system. For this exercise, assume a platform runs long-lived AI agent tasks for many organizations. Each organization receives monthly token and tool-call allowances, may divide limits among projects, and may run concurrent agents.
Before an agent step starts, the platform must reserve an enforceable upper bound; the runtime stops the step at that bound, then records actual metered usage and releases only proven-unused capacity.
A retry, timeout, or duplicated meter event must not charge twice, and accepted reservations must not let a tenant exceed its hard quota. Administrators can inspect usage, change future limits, and reconcile disputed totals. Model-provider billing and the internals of the agent runtime are outside scope.
What to build
- An agent runtime can idempotently reserve a hard token and tool-call ceiling before starting a step.
- The platform can settle usage within that ceiling, release only proven-unused reservation, and deduplicate repeated meter events.
- Tenant and project administrators can inspect available, reserved, consumed, and reconciled quota.
- Authorized administrators can version future quota allocations without rewriting historical usage.
At scale
- For this exercise, assume 100,000 tenants, 10 million active agents, and peaks of 200,000 reservation or settlement commands per second.
- For this exercise, target 100 milliseconds p95 for a reservation decision in a healthy region and make settled usage visible within one minute.
- For this exercise, quota is a hard safety boundary: retries may be rejected conservatively, but the system must not knowingly overspend a tenant limit.
- For this exercise, retain an auditable usage ledger for 13 months and survive a single-zone failure without losing acknowledged reservations.