Don't reach for swarm until 50+ agents — and decide partial-failure modes at design time

Digital Applied·medium signal

For 3–10 concurrent agents, a supervisor with fan-out branches is simpler, cheaper, and more debuggable than a swarm; swarm's infrastructure overhead only pays off above ~50 concurrent agents. The harder discipline is fan-out failure handling: specify up front whether a single failed branch fails the whole request, returns partial results, or retries only that branch — leaving it implicit is what turns into production incidents.