What measurement approach should I use to model the incremental infrastructure cost of durable execution per agent step?

Instrument a representative workload to capture CPU, memory, network, and storage overhead per step. Compare that against a baseline of equivalent non-durable orchestration. Use the differential to build a capacity model that accounts for state serialization, journal writes, and replay traffic. A boundary: overhead varies with step payload size and workflow topology, so results from synthetic benchmarks may not transfer directly to production patterns.