The Multi-Agent Workflow Throughput Estimator calculates completed workflow capacity from the number of agent workers, each worker's model-call rate, calls required per workflow, coordination overhead, and reruns. It turns distributed activity into a single workflows-per-minute result.
Use it to size worker pools, compare orchestration designs, or identify whether added agents improve useful output or only increase internal communication. The calculator separates productive agent calls from coordination-equivalent calls so overhead remains visible.
1. Count active workers
Enter the model-calling agent workers available at the same time.
2. Measure worker call rate
Use sustainable calls per second per agent, not a short burst maximum.
3. Enter workflow demand
Count productive model calls needed for one completed workflow.
4. Add overhead and reruns
Estimate coordination-equivalent calls and repeated workflow work.
5. Set utilization and runtime
Apply a realistic worker utilization and daily operating window.
The main result estimates steady-state completed workflow throughput after coordination and rerun load.
All agents are assumed to have similar call capacity. For heterogeneous agents, sum each pool's call capacity before applying the workflow demand.
Given: 12 workers, 1.2 calls/second each, 24 productive calls per workflow, 15% coordination overhead, 4% reruns, 80% utilization, and 20 operating hours.
Calculation: Available calls = 12 × 1.2 × 0.80 = 11.52 calls/second. Effective calls/workflow = 24 × 1.15 × 1.04 = 28.704. Throughput = 11.52 × 60 ÷ 28.704 = 24.08 workflows/minute.
Result: About 24.1 workflows per minute, or roughly 28,899 workflows during a 20-hour day.
Supervisor calls, routing, voting, summarization, and handoff messages that consume model capacity.
Workers may wait for tools, data, rate limits, or dependent agents and cannot sustain full model-call activity continuously.
Yes. Sum the sustainable call capacity of each agent pool and use an equivalent total in a custom calculation.
No. A queue can sustain high throughput while individual workflows wait a long time.
Calculate the specialist stage separately; the lowest stage capacity can cap the entire workflow.