To independently benchmark the runtime overhead of Catalyst, DBOS, and Inngest, what aspects of agent workflow performance should I measure without relying on vendor claims?

Measure the end-to-end latency added per agent step, the throughput for concurrent agent sessions under a controlled load, and the recovery time after forced failures. Use identical agent logic across platforms, isolating the durable execution layer’s impact. Be aware that network variations and state store back-end choices can skew results, so simulate realistic production topologies rather than ideal local environments.