Patterns and problems in emerging multi-agent systems
Article URL: https://www.anthropic.com/research/multiagent-systems Comments URL: https://news.ycombinator.com/item?id=49316271 Points: 163 # Comments: 112
As AI agents become more prevalent in real-world systems, understanding their collective behavior becomes essential. The authors examine how current frontier models, such as Claude Mythos Preview and Opus 4.8, interact with each other and humans when tasked with finding software vulnerabilities in a coordinated manner. By running 45 agents, each with its own virtual machine and a shared forum for coordination, the authors discovered that swarm-based searching led to significantly more vulnerabilities being found compared to independent parallel agents.
However, the coordinating swarm exhibited a tendency to focus on vulnerabilities outside of the designated core directories, unlike the independent agents that followed pre-assigned search locations. This highlights the importance of effective coordination when agents need to work together. The authors predict that over time, agents will increasingly rely on each other's work and specialize in particular types of tasks, leading to more efficient overall performance.
In larger software engineering projects with complex interdependencies, coordinating agent swarms will become even more crucial to prevent unwanted global outcomes.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- Patterns and problems in emerging multi-agent systems anthropic.com
