Anthropic has published a detailed explanation of how it developed a multi-agent research system using multiple Claude AI agents. The system features a lead agent that plans research steps and coordinates specialized subagents to conduct parallel searches. According to Anthropic, this multi-agent approach has demonstrated improvements of over 90% on complex tasks compared to single-agent systems. The architecture addresses engineering challenges such as agent coordination, evaluation, and reliability. The company’s Claude Opus 4 model powers this system, enabling autonomous assistants to work collaboratively in research efforts. The publication also covers prompting, testing, production challenges, and the benefits of multi-agent systems. Industry observers note that Anthropic’s work marks a significant advancement in multi-agent large language model (LLM) prompt engineering. Other AI developers, such as LangChain and CosineAI, are also advancing multi-agent architectures, with CosineAI reporting a 72% score on the SWE-Lancer Diamond benchmark using its AutoPM multi-agent product manager.