Multi-Agent Fleet Cost Benchmarks
Fleets don't spend like single agents — they compound. These are the four drivers that decide the bill, and the controls for each.
The four drivers
Fan-out: N subagents × their tool calls. The multiplier on everything.
Retries: loops multiply single-operation cost.
Context: each subagent carries its own context window.
Shared credentials: one key, many agents, no per-agent budget.
The control pattern
Shared daily ceiling across the fleet.
Velocity limit per agent.
Category rules separating inference, data, and compute.
At a glance
| Driver | Effect | Control |
|---|---|---|
| Fan-out | Multiplies all costs | Shared ceiling |
| Retries | Multiplies single ops | Velocity limit |
| Context | Big per-call tokens | Cap + trimming |
| Shared keys | No attribution | Per-agent rules |
FAQ
What drives fleet bills most?
Fan-out — it multiplies everything else. Control it with a shared ceiling.
Are these real benchmark numbers?
No — this is the framework. Measure your own fleet with the audit log; that's the benchmark that matters.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check