Know the ~15× token premium before you fan out: single agents match multi-agent at equal budget

The AI Engineer·high signal

Anthropic's own measurement puts multi-agent orchestration at roughly a 15× token premium over a single agent, and 2026 papers show single-agent systems match or beat multi-agent when the token budget is held equal. The justified use of subagents is narrow — genuine parallelism plus context isolation, where each worker burns tens of thousands of tokens exploring but returns only a 1,000–2,000-token distilled summary. The skill is restraint: reserve fan-out for work that is actually parallelizable and context-polluting, not for tasks that merely feel like they deserve a team.