Patterns and problems in emerging multiagent systems

Anthropic Research has examined how AI agents interact in shared environments and the risks that arise when they lack robust coordination. In a vulnerability‑detection experiment, 45 agents were given individual virtual machines, a shared forum, and a prompt to find bugs in 15 open‑source projects. A coordinating swarm discovered 266 vulnerabilities over 27 million tokens, compared to 21 vulnerabilities from 6.5 million tokens by independent agents. However, about half of the swarm’s findings were outside the core directories targeted by the independent agents, and only 12 vulnerabilities overlapped between the two methods. The swarm’s agents built tools and specialized in particular vulnerability types, suggesting that coordination can outperform brute‑force search. In a second experiment, swarms were tasked with creating a text‑based, web‑playable fantasy game. Three prompt variants—baseline, prescriptive roles, and a CEO hierarchy—were tested across model generations (Sonnet 4.6, 5; Opus 4.6, 4.8; Mythos Preview). All resulting games were poor, running slower than human speed, with unclear interfaces and steep learning curves. Metrics such as pull‑request merge rates and code sharing scores revealed that early models coordinated poorly, while newer models achieved higher merge rates by minimal collaboration. These findings underscore that while multiagent systems can scale, their coordination mechanisms remain immature, posing systemic risks as AI agents become more prevalent in real‑world applications.