Anthropic details Claude multi-agent research system.

This is a strong proof point that multi-agent design is becoming practical in production, but Anthropic is also clear that orchestration quality and cost control are still the hard parts.

• –Anthropic says its multi-agent stack outperformed single-agent Claude Opus 4 by 90.2% on internal research evals, showing clear upside for parallel exploration tasks.
• –The architecture uses a lead planner plus specialized subagents, then a citation stage, which mirrors how many teams are now designing agent workflows in enterprise tooling.
• –The post emphasizes that prompt design, tool descriptions, and eval loops mattered as much as model choice, which is a key takeaway for developer teams building agents.
• –Token cost is a real tradeoff: Anthropic reports multi-agent research runs can consume far more tokens than standard chat, so this pattern fits higher-value tasks best.
• –Anthropic explicitly calls out coding as a weaker fit for multi-agent parallelism right now, reinforcing that “more agents” is not a universal optimization.