Cost per completed multi-agent task should include unsuccessful runs, retries, specialist calls, and recovery effort. Counting only the final agent’s usage understates what the architecture spends to produce a correct result.
Unit cost needs a verified denominator
Multi-agent task unit cost divides attributable included cost by correctly completed logical tasks over a stated cohort. Keep task identity separate from agent messages and execution attempts. A task with five handoffs is still one business task.
For a proposed cost ledger, include model usage, tools, retrieval, infrastructure, external effects, and relevant staff review. Separate observed billing from estimates and describe delayed or missing usage data.
Define completion through authoritative records or the approved evidence rubric. A final message saying “done” does not establish the denominator. Safe fallback and unresolved work should remain visible as separate outcomes.
Calculate a fictional architecture comparison
A single-agent design costs 100 units across 50 evaluated tasks and correctly completes 40. Included cost per correct task is 2.5 units. A specialist design costs 180 across the same 50 tasks and correctly completes 45, yielding four units per correct task.
The specialist design completes more tasks but costs more per correct outcome. Whether that tradeoff is worthwhile depends on the business consequences and supported scope, not a universal preference. The figures are illustrative, not measured benchmarks.
| Cost component | Attribution rule |
|---|---|
| Specialist calls | Parent logical task reference |
| Retry attempts | Same task, included consumed resources |
| Shared retrieval cache | Explicit allocation method |
| Failed tasks | Included numerator cost |
| Staff recovery | Measured or labelled estimated effort |
Compare equivalent scenario mix and configurations. A harder cohort assigned only to one architecture makes a direct cost ranking misleading.
Keep critical quality failures outside averages
A lower unit cost cannot compensate for unauthorised writes or false commitments under a blocking policy. Show those defects beside cost and completion. Latency and staff burden are separate operational dimensions worth reviewing.
Precise allocation can be difficult when providers bill in aggregate or shared services serve several tasks. State the method and uncertainty, then reconcile with invoices. Avoid using a decimal-heavy number to imply accuracy the data does not support.
This cost model is proposed. The voice integration case study provides related engineering context without claiming comparative model costs. Pair it with single-agent architecture choice.
Multi-agent unit cost FAQ
Should failed tasks be excluded from spend?
No. Their consumed resources belong in the included numerator.
Is cost per agent message useful?
It can diagnose usage, but it does not measure the cost of a correct completed business task.
What should an architecture decision compare?
Use outcome quality, critical defects, end-to-end latency, total included cost, and operational effort under comparable fixtures.