Autonomous Agents: How to Set 3 Spend Limits They Cannot Edit

Spend limits for autonomous agents must be set by the developer tool administrator and cannot be edited by the agent itself. This ensures the invoice remains within a defined ceiling, preventing overcharges. There are three key boundaries: policy ownership, reservation before execution, and reconciliation against measured usage.

Each boundary has its own failure mode. The agent can request work and observe its remaining allowance but cannot change the allowance itself. Separate credentials and least privilege access control are essential for different roles (policy owner vs. agent). The ledger must maintain the invariant: committed + reserved = ceiling for each customer and billing period.

A request has a unique idempotency key across retries, which links its reservation, consumption, release, and invoice entry. An example scenario with two jobs for customer-42 shows how the reservation system works. The units in the system matter; a gateway must reserve a defensible upper bound and enforce a per-call maximum at the executor.

If a job cannot be bounded or interrupted, a strict ceiling cannot be guaranteed by admission control alone. The policy decision on admission (hard reservation or soft alert) depends on the customer's requirements rather than the agent's optimization goals. The Go example provided illustrates the admission rule, which should be implemented in production storage with serialization, policy change authorization, and a settlement trail.