An agent acts. A swarm acts in parallel. But only for jobs that don't wait on each other.
I kept being the slow part. The AI could do a focused job end to end, but I was handing them out one at a time and watching each finish. The unlock was naming the three sizes of the same thing: an agent (an AI handed tools and let loose on a job), a subagent (one dispatched for a single focused task in its own clean context), and a swarm (many at once, in parallel).
I tested swarming — launching independent jobs together instead of in sequence. A tiny decision script answers the only question that matters: do these jobs depend on each other? No → swarm them. Yes → sequence them.
Swarming independent work. Six two-minute jobs land in about the slowest one's time, each in its own context, instead of twelve minutes nose-to-tail.
Swarming dependent work. Launch a job that needs another's result and you get an agent confidently building on something that doesn't exist yet. The edge is sharp — independence is the whole call.
You're the foreman, not the bricklayer. Swarm what's independent; sequence what isn't.