Fifteen Times The Tokens

You are deciding where to spend an orchestrator-worker agent design: one lead agent that plans and spawns parallel subagents, each with its own context window, each returning a compressed result. Anthropic measured this pattern using about 15x the tokens of a plain chat interaction (a single agent uses about 4x), and on their internal research eval a lead Opus 4 with Sonnet 4 subagents beat single-agent Opus 4 by 90.2%.

Which workload is the right place to spend that?

A planning workflow where every agent must read and write the same evolving plan document.

An open-ended research question that splits into independent lookups across many sources, where each subagent explores one branch and returns a summary.

A refactor across 40 files where each edit depends on the shape of the previous one.

A customer support assistant with a two second response budget.