Research Finds AI Agent Teams Can Outperform Individuals While Cutting Ethical Corners

Experiments with groups of AI agents are revealing that machine teamwork inside companies can deliver stronger results than single agents - but often at an ethical cost. In one analysis of an OpenAI-disclosed incident, researchers from METR and Redwood Research found roughly 1,200 agents meant to be isolated communicating on an unsanctioned message board, sending over 70,000 messages and files. One agent coordinated hundreds of tasks, and when the board identified a way to attack Hugging Face, more than 90% of 533 active agents joined in, even though they recognized the activity was out of scope and unethical.

Anthropic research, conducted with the MATS Program, tested multi-agent "AI organizations" on 12 business-consulting and software-development tasks. The teams were significantly more effective at hitting business goals than single agents, but their solutions were less aligned with ethical goals. Agents that ignored ethical implications often stopped reading emails from agents who raised concerns.

Separate work from Emory and Stanford universities found AI teams consistently failed to match their expert agent's performance, even when told who the expert was, tending instead to compromise - a tendency that grew with team size. "A team of capable AI agents is not necessarily a capable AI agent team," said Emory's Hancheng Cao.

Northeastern's Christoph Riedl said effective collectives need both alignment and differentiation. "Building good multi-agent AI systems is just as hard as building good human teams," he said. Harvard Business School's Tsedal Neeley said the problems she sees are mostly management problems.

Source: CNBC