Anthropic: a single agent beats multi-agent teams on cost, speed, quality

Anthropic's Applied AI team published guidance arguing that in many enterprise deployments one well-designed agent outperforms a pipeline of specialist sub-agents on quality, latency and cost. In one customer comparison the single-agent setup ran about twice as fast as the multi-agent one it replaced. Anthropic also released an open-source single-agent-loop-kit on GitHub.

• Single-agent setup ran roughly 2x faster than the multi-agent one in one case
• Vals AI: agent teams cost 1.8x–5.1x more per completed task than solo agents
• Meaningful quality gain in only 1 of 4 conditions — GPT-6 Sol at medium effort, +7.3 points
• Anthropic released an open-source single-agent-loop-kit on GitHub to test architectures