CH 02 · The Case For Multi-Agent: Anthropic's Research System · 02/08
Read as one pageWhy research needs more than a single agent
The motivation is the read-versus-write idea from Chapter 1. Open-ended research does not follow a predictable path. Each discovery can shift the direction of inquiry, so you cannot script a fixed pipeline in advance. A single agent also runs into two hard walls: the context window fills up, and its sequential processing makes broad searches slow.
The articles also contrast this with traditional RAG. Standard RAG fetches a fixed set of documents once, based on similarity to the query, and then answers from them. Research needs more: it needs to search repeatedly, adapt based on what it finds, and follow new leads. A multi-agent system can do this dynamic, branching exploration in a way a single static retrieval cannot.
The payoff was large. In internal evaluations, a system using Claude Opus 4 as the lead agent and Claude Sonnet 4 as subagents outperformed a single-agent Claude Opus 4 by 90.2 percent on their research evaluation. The articles are candid that this works largely because multi-agent systems spend more tokens: token usage alone explained about 80 percent of the performance variance in their browsing evaluations, with the number of tool calls and the model choice as the next biggest factors.
Finished this idea?
Your learning progress stays in this browser.