How should I define service objectives for agent workloads with probabilistic steps?

A recommended first step for defining service objectives for agent workloads with probabilistic steps is separating deterministic and probabilistic workstreams. Next, map clear, specific success criteria to each deterministic workstream first, then set key outcome-focused targets for probabilistic phases such as model inference runs. It is important to note that durable execution does not provide exactly-once delivery, so avoid tying core service objectives to single model inference attempts on their own.