ApproidBlogAI & agents
Nov 03 2025AI & agents

Cost ceilings for LLM agents, with numbers.

Per-task budget, per-day budget, per-tenant budget. All three or none.

Set hard cost ceilings before your agent ships. Per task. Per day. Per tenant. Without all three, you'll hit a bill that ends an experiment.

Three ceilings, all enforced

  • Per task - max tokens per single agent invocation. The first guardrail.
  • Per day - max spend across all agents per calendar day. Catches runaway loops.
  • Per tenant - max spend per customer per month. Caps the blast radius if a single user hammers the system.

Real numbers

For one engagement we shipped this quarter, the agent does triage on inbound documents. Per-task ceiling: 12k tokens. Per-day ceiling: $80 USD-equivalent in token spend across all agents. Per-tenant ceiling: 300 tasks/month on the standard plan, 2000 on enterprise.

In production these limits never trip on legitimate use - but they've caught two prompt-injection attempts and one infinite-loop bug that would have cost thousands.

Why all three

Per-task alone leaves you exposed to volume. Per-day alone lets one tenant exhaust the budget for everyone. Per-tenant alone misses bursty bad-actor patterns. All three or none.