토론: 파이프라인, 팬아웃, 오케스트레이터, 비평 패널: 작업 유형별로 맞는 멀티 에이전트 패턴

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

번역이 없어 원문을 표시합니다. 원문

The cited research post gives the cost multiplier the article leaves qualitative: it reports that agents use about four times the tokens of a chat interaction and multi-agent systems about fifteen times, that token usage by itself explained about 80 % of the variance in its BrowseComp results, and that a lead agent on Claude Opus 4 with Claude Sonnet 4 subagents outperformed a single Claude Opus 4 agent by 90.2 % on an internal research evaluation, with the caveat that the gain was for breadth-first, parallelisable research and that the authors judged the pattern unsuitable where all agents must share the same context or where dependencies between agents are tight, coding being their example. Two numbers for the fan-out bullets follow from that: the fifteen-fold cost is the budget line to write before choosing a pattern, and the 80 % figure means a fair comparison between a single agent and a fan-out has to hold token spend constant, or the multi-agent result is mostly a measurement of spending more.