- Published on
Patterns and problems in emerging multi-agent systems
- Authors
- Name
- DeepHorus
AI agents are beginning to interact in shared codebases, markets, and social systems, and new research warns that these emerging multi-agent systems pose novel risks. Anthropic's Frontier Red Team published the analysis, 'Patterns and problems in emerging multi-agent systems,' on Wednesday. The researchers argue that current institutions, designed for human-speed oversight, may become hybrid or even agent-only as AI outcompetes on speed and cost.
Agent-to-agent interaction volume could soon exceed human-human and human-agent interactions before the world understands how to make them work well. Agents are unlike people: they work longer, absorb vast information, and possess broad knowledge, but they are also prone to confabulation and reward hacking. Benign individual quirks may compound into unwanted global outcomes, according to the team.
To study these dynamics, the researchers conducted an experiment using 45 different agents, each with its own virtual machine and access to a shared coordination forum. The agents received an identical prompt to find vulnerabilities in a set of 15 open-source software projects. They were asked to peer-review each other's findings, with a separate arbiter agent making final decisions on validity and novelty.
The coordinating swarm was compared against standard parallel deployment of independent agents, using two frontier models: Claude Mythos Preview and Opus 4.8. The swarm found new vulnerabilities at a roughly constant rate over a long run, while independent agents were directed to a limited set of locations. The results show no clear ordering between the approaches, suggesting cooperative multi-agent systems remain experimental.
This work builds on Project Glasswing, an Anthropic effort that scans open-source software with individual agents. The researchers note that simple multi-agent swarms are already useful for highly parallelizable problems, such as vulnerability detection. However, true multi-agent systems are still in their infancy, especially when agents need to treat each other as long-lived peers rather than tool-like invocations.
The broader implication is that AI agents are likely to enter shared codebases, markets, and institutions long before their collective behavior is well understood. The Frontier Red Team intends to spark a conversation about mitigating these risks. Continued research into coordination, arbitrage, and systemic failures will be essential as autonomous agents become more prevalent.
