Anthropic: Multi-Agent Collaboration May Be the Next Trend in AI, but Complex Tasks Still Face Significant Coordination Challenges

Odaily News: Anthropic published a research article on emerging multi-agent systems, noting that as AI Agent capabilities improve, the scale of interactions between Agents may exceed that of human-to-human and human-to-Agent interactions in the future. However, current models still face coordination difficulties in complex collaborative environments.

In a software vulnerability detection experiment, Anthropic deployed 45 Agents simultaneously. Claude Mythos Preview's independent parallel mode consumed 6.5 million tokens and discovered 21 vulnerabilities, while the collaborative Agent cluster consumed 27 million tokens and discovered 266 vulnerabilities in total. Anthropic stated that during collaboration, Agents autonomously develop tools and gradually form specialized divisions of labor. It is expected that this "specialization + collaboration" model will outperform purely parallel brute-force search in the future.

However, in software development tasks requiring high interdependence, multi-agent performance remains poor. Anthropic tested various organizational approaches, including free collaboration, predefined roles, and the setup of a "CEO Agent," but none significantly improved the final outcomes. At this stage, complex multi-agent projects still require substantial human guidance.