The Problems Scale Exposes
Everyone's tired of the single-agent story; the value of this piece is in the order of magnitude of "hundreds": when hundreds of agents work in parallel, new problems crop up layer by layer. The cost curve is far steeper than expected, and failed retries burn money like water at scale; work among agents needs coordination and de-duplication, or half the compute goes to duplicated labor; and aggregating and arbitrating results becomes the new bottleneck—among a hundred "I'm done" reports, how many are false claims requires a dedicated verification layer to check. The author's record is honest, sparing neither the crashes nor the gains.
Reference Value for Those Who Follow
Multi-agent is one of the hottest directions right now, but public experience is mostly demo-grade, and an engineering field report like this—with cost figures and failure cases—is scarce. Copiable conclusions include: build the verification mechanism before scaling, because without a reliable "acceptance inspector" an agent army is just a noise amplifier; assess a task's parallelizability in advance, because piling agents onto a strong dependency chain is a waste; and hardcode budget guardrails, because agents won't spare your money on their own. The author's quip lands well too—managing a hundred agents is like managing a hundred overconfident interns; the test is management, not prompting.
via: Hacker News