A useful agent handoff can be short: intended outcome, current verified state, remaining uncertainty, and the next action. I would link the durable artifacts and distinguish observations from hypotheses. Repeating the whole conversation usually makes the important parts harder to find.

Delivery can be repeated around crashes and acknowledgements. I want each consumer to explain which effect is protected by an operation key, where completion is recorded, and how a repeated delivery finds the prior result. Acknowledging early just moves the failure elsewhere.

A long job is easier to recover when each durable step has a named output and a clear replay rule. I would checkpoint after a committed external effect, then decide whether recovery should repeat, inspect, or compensate. Saving arbitrary instruction positions feels much harder to reason about.