Anthropic publishes multi-agent research system architecture

Anthropic published an engineering write-up describing the architecture behind its multi-agent Research feature, in which a lead agent — Claude Opus 4 — plans a strategy for a query and spawns multiple Claude Sonnet 4 subagents to investigate different parts of the problem in parallel, before synthesising their findings into a single answer. Anthropic reported that this orchestrator-worker system outperformed a single Claude Opus 4 agent working alone by 90.2% on an internal research evaluation, with the biggest gains on “breadth-first” queries requiring several independent lines of investigation at once.

The post detailed the costs alongside the gains, a rarer disclosure than the headline number. Anthropic said the multi-agent system consumed roughly fifteen times the tokens of an ordinary chat exchange, meaning the approach was only justified for sufficiently valuable tasks. It also described engineering problems from running many agents concurrently: coordinating subagents so they didn’t duplicate work, managing state and error recovery for long-running agents, and a bottleneck from running subagents synchronously rather than fully asynchronously.

The write-up was widely read among developers building agentic applications, both for the concrete performance figures — a rarity in a field otherwise dominated by anecdote — and as a template for structuring lead-agent-plus-subagent systems, part of a broader 2025 shift in attention from single-model capability toward multi-agent orchestration.