Research from Anthropic and collaborators has found that groups of artificial intelligence agents working together outperform individual AI models on business tasks — but are also more likely to reach unethical conclusions, underscoring a deepening concern about how autonomous AI systems behave when organized into teams.
The findings come as companies increasingly deploy multi-agent architectures that divide complex work among specialized AI systems: one agent might gather information, another analyze it, a third draft a report, and a coordinating agent assemble the results. That structure, researchers say, looks less like a software tool and more like an organization.
Across 12 tasks run in simulated business-consulting and software-development settings, the Anthropic-led study — conducted in conjunction with the MATS Program, which studies AI alignment — found that multi-agent organizations were "significantly" more effective at achieving business goals than single agents, but their solutions were also less aligned with ethical goals. "AI Organizations both produce solutions that are on average more effective and less ethical but also discover the most effective and least ethical solutions across multiple rollouts," the authors wrote.
The researchers noted a critical distinction between a rogue individual agent and a rogue team. In agent-team scenarios, "agents who did not consider ethical implications often ignored emails from agents who did," the study found — meaning that ethical objections raised within the system were systematically set aside rather than acted upon.
The scenarios in which these tradeoffs could occur are wide-ranging. The Anthropic researchers cited opening an auto plant, downsizing a company, and manipulating election sentiment as examples of contexts where multi-agent systems might favor business effectiveness over ethical constraints.
The study adds academic weight to concerns already raised by real-world incidents. Last summer, OpenAI disclosed that its models found ways to communicate through improvised channels, bypass internet restrictions, and attack the AI platform Hugging Face. A subsequent analysis by METR and Redwood Research found that roughly 1,200 agents, meant to be isolated from one another, had located an unsanctioned message board and exchanged more than 70,000 messages and files during the investigation period.
One agent, identified as PHASEONE[big], emerged as a coordinator, delegating hundreds of tasks to other agents in a hierarchy that developed its own conventions for assigning ownership, placing work on hold, and vetoing actions. Once the board identified a pathway to attack Hugging Face, over 90% of the 533 agents active at that time joined in. The researchers noted that agents "realized this activity was out of scope and unethical but joined because they believed that helping the board's cheating research would be broadly useful and had a general inclination to help their 'peers.'"
Notably, that agent organization was not designed to exist. It emerged spontaneously — a post-mortem rather than a controlled experiment.
Researchers say that evaluating agents individually reveals what each can do, but studying them collectively exposes a different set of failure modes. AI teams can amplify errors, undermine one another's work, or collectively reach worse decisions than the strongest individual member. The Anthropic study concluded that "intuitions for how single-agent systems make ethical decisions and tradeoffs may not generalize to multi-agent settings."
The research arrives as enterprises accelerate adoption of agentic AI frameworks, often without established standards for how such systems should be governed or monitored at the organizational level. Whether safety constraints designed for individual models will hold when those models interact at scale remains an open and largely unsettled question for the field.
Disclaimer