Anthropic published a research article on emerging multi-agent systems, saying that as AI agent capabilities improve, interactions between agents may eventually exceed interactions between people and between people and agents. According to Odaily, the company said current models still face coordination difficulties in complex collaborative environments.

In a software vulnerability detection experiment, Anthropic deployed 45 agents at the same time. The independent parallel mode of Claude Mythos Preview used 6.5 million tokens to find 21 vulnerabilities, while a collaborative agent cluster used 27 million tokens to find 266 vulnerabilities. Anthropic said agents developed tools on their own during collaboration and gradually formed specialized roles, and it expects a future model of specialization plus collaboration to outperform simple parallel brute-force search.

Anthropic also said multi-agent systems performed poorly in software development tasks that require high interdependence. The company tested free collaboration, preset roles, and organizational structures such as a CEO agent, but none significantly improved the final results. It said complex multi-agent projects still require substantial human guidance at this stage.