Research agents buy breadth with a much larger token budget

The research system described here uses a lead agent to divide a question among workers and combine their findings. That arrangement works especially well when useful evidence can be gathered along several independent paths.

Anthropic reported a large gain on its internal research evaluation, but also said its multi-agent system typically used about fifteen times as many tokens as a normal chat. The comparison therefore combines a different organization of work with substantially more computation.

The company explicitly identifies limits: work requiring tightly shared context or many dependent steps is harder to divide. More agents can also mean redundant searching, coordination failures, and larger costs. The account is useful as a design study of broad research tasks, not a general demonstration that adding workers improves every kind of knowledge work.