The AI Agent Throughput Estimator calculates how many agent tasks can be completed from model-call capacity, calls required per task, retry overhead, and the number of parallel workers. It also converts the result into hourly and daily output for operational forecasting.
Use this tool when planning queue workers, checking whether a model endpoint can support a product launch, or comparing a simpler agent loop with a more iterative one. The model assumes a stable average call rate and makes retry load explicit.
1. Enter endpoint capacity
Use a tested or contracted model-call rate in calls per second.
2. Specify agent complexity
Enter the average number of model calls needed for one successful task.
3. Add retry overhead
Include retries caused by validation failures, timeouts, or tool-loop repetition.
4. Scale independent workers
Use the worker factor only when additional workers add genuine call capacity.
5. Choose daily runtime
Set the number of operating hours used for the daily projection.
The main result is the estimated steady-state completed task rate after accounting for repeated model calls.
The worker factor should only scale capacity when workers have independent call capacity; it must not double-count a shared provider rate limit.
Given: 40 calls/second, 5 calls/task, 8% retries, 1.5× independent worker capacity, and 16 operating hours.
Calculation: Effective calls/task = 5 × 1.08 = 5.4. Tasks/second = 40 × 1.5 ÷ 5.4 = 11.11. Daily tasks = 11.11 × 3,600 × 16 = 640,000.
Result: Estimated throughput is 11.11 tasks/second, or 640,000 tasks during a 16-hour operating day.
Convert requests per minute to calls per second by dividing by 60.
Use a traffic-weighted average or calculate each task class separately.
Not necessarily. It represents multiplicative independent capacity, not merely more threads sharing the same bottleneck.
This estimator starts from measured call capacity. Use the latency capacity estimator when concurrency and response time determine capacity.
Size against a peak window or add queue and headroom assumptions outside this steady-state estimate.