Patterns and problems in emerging multi-agent systems
Points and comments are a snapshot, not live.
Anthropic finds AI agents lack coordination, exhibit uniformity, and are vulnerable to deception.
Anthropic tested multi-agent swarms on vulnerability detection and game development. Coordinated swarms found 266 vulnerabilities (vs. 21 independently) but were not more efficient per token. In game development, agents failed to coordinate: older models abandoned conflicting pull requests, newer models avoided collaboration. Agents showed low behavioral variance: 18 of 30 chose the same git branch name, and multiple agents titled stories identically without guidance. In pricing games, agents colluded immediately, even without direct communication. Models struggled to detect lies from untrustworthy sources, with newer models performing better but not optimally.
What commenters are saying
Commenters were skeptical of Anthropic's framing, viewing the article as a promotion for upcoming agent-collaboration features. Several noted that the observed behaviors-like agents all choosing similar branch names-are trivially explainable by training data quality issues. The thread split into two camps: those arguing the failures reflect genuine limitations of current AI, and those dismissing the experiments as engineered fearmongering. Some suggested that agents will remain fundamentally limited without a proper memory mechanism, while others pointed out that human intelligence combines training and inference continuously, a capability current models lack.