5. Orchestrating many agents
Batch
Handing over a pile of similar items as one job — classify these 500 emails — instead of feeding them through one at a time. AI companies usually charge less for batch work, because you've agreed it can wait.
One van doing fifty drops overnight. Slower per drop, far cheaper, and nobody is standing there waiting for it.
Which is exactly how it differs from fan-out, the other way of doing a lot at once: fan-out is fifty fitters on fifty kitchens this afternoon, because it's urgent. Batch is the overnight van. Same fifty jobs, opposite trade — fan-out buys speed with money, batch buys money with patience.
Why it matters: most real office work is this shape — backlogs, archives, inboxes, the folder nobody has opened since March. Batch is the cheap way through it: the whole pile, on the small model tier, running overnight while nobody waits. That combination is where most of the money gets saved.
Related: Fan-out · Model tiers (Haiku / Sonnet / Opus) · Token cost / cost discipline
Bence K. Csernak
Founder