A rollback that restores the old configuration may still leave the system broken.
State does not roll back because the file did.
I have watched routine changes cause disproportionate incidents: a firewall rule, DNS shift, certificate renewal, or failover that looked harmless in isolation.
The configuration was reverted, but the service did not recover. Sessions had been dropped. Caches were stale. Routing state had diverged. A downstream system had already accepted part of the change. Agentic workflows add the same problem at another layer: tool calls, retries, durable tasks, and external actions can leave partial state behind even when the agent itself stops.
Rollback design has to identify what is restored, invalidated, replayed, or compensated.
"Previous version available" and "operation reversible" are not the same property.