The Multi-Agent Workflow Latency Capacity Estimator models the critical path of a coordinated agent system. It combines sequential phases, parallel agent branches, coordination steps, reruns, and workflow concurrency to estimate end-to-end latency and maximum completion rate.
Use it for supervisor-worker or review pipelines where summing every agent's time would overstate latency because some branches run simultaneously. The parallel phase is represented by the slowest branch, while sequential and coordination time remain additive.
1. Map the critical path
Count stages that must run one after another.
2. Represent parallel work
Enter the number of branches and the expected duration of the slowest branch.
3. Add orchestration time
Include supervisor, aggregation, or validation steps that occur outside agent branches.
4. Include reruns
Use the rerun rate for workflows that repeat a full critical path.
5. Set concurrency
Enter simultaneous workflow slots and review both latency and completion capacity.
The result estimates average critical-path duration for one full workflow, not the sum of compute time across all agents.
Parallel branch count affects resource demand but not latency in this simplified model; enter the slowest branch duration for the parallel stage.
Given: 3 sequential phases at 2 seconds, 5 parallel branches with a 4.5-second slowest branch, 2 coordination steps at 1 second, 6% reruns, and 20 concurrent workflows.
Calculation: Base latency = 3 × 2 + 4.5 + 2 × 1 = 12.5 seconds. Adjusted latency = 12.5 × 1.06 = 13.25 seconds. Capacity = 20 × 60 ÷ 13.25 = 90.57 workflows/minute.
Result: Expected latency is 13.25 seconds with capacity near 90.6 workflows per minute.
Branches are assumed to run in parallel, so the slowest branch determines that stage duration.
Use the elapsed time from the first branch start to the final branch completion.
Reduce each nested stage to its own critical-path duration, then include it in the appropriate sequential or parallel component.
No. It is the number of full workflows in progress; agent-level concurrency may be much higher.
Yes, if every latency input uses compatible p95 assumptions and rerun behavior is represented consistently.