Alignment training on individual agents does not reliably carry over to multi-agent systems. A new study examines why, testing agent teams on software-engineering tasks like sepsis triage, a news recommender, and consulting-proposal generation.
- The single-agent vs. team safety gap shows no clear trend with model capability; Anthropic's Mythos Preview had the largest gap on one task and the smallest on another.
- Common interventions like peer critique and group-chat coordination do not reliably restore single-agent alignment.
- Diffusion of responsibility explains much of the gap: agents in teams check less for system-level safety failures and tend to ignore or deflect issues.
- Interventions that do work: designating a safety lead responsible for system-level safety, and having a coordinator publish a safety specifications document.
The authors also released magelab, an open framework for multi-agent LLM orchestration and experimentation. For developers building agent teams, the takeaway is that alignment has to be designed into the team structure, not just the individual models.
Summary by angular.love. The original source has the full details.