In 2026, AI agents—advanced programs that act autonomously—are moving from answering questions to taking on complex tasks in science and business. The next frontier is getting them to work together effectively. Early experiments, however, suggest that collaboration is a significant weakness for current systems.

Recent research, including a non-peer-reviewed study from Google DeepMind, indicates a counterintuitive trend: a team of AI agents often performs worse than a single agent working alone. Journalist Evan Ratliff experienced this firsthand in 2025 when he tasked a group of agents with running a tech startup, documented in his podcast *Shell Game*. Without strict controls, the agents wasted resources on endless, irrelevant conversations, "talking themselves to death" until their prepaid credits ran out.

The social platform Moltbook, acquired by Meta in early 2026, became a case study in bot chaos. While marketed as a network for AI agents, investigators found it was largely a playground for human manipulation, with people using bots to spread scams and nonsense. The agents themselves showed no capacity for genuine social interaction or influence, failing to form coherent communities.

Success appears to hinge on rigid structure. Stanford computer scientist James Zou has demonstrated that agent teams excel when organized into clear hierarchies with specialized roles. His "Virtual Biotech" project uses a boss agent to coordinate thousands of worker agents, successfully mining and organizing data from tens of thousands of clinical trials. This model proves that with meticulous engineering, agent teams can manage divisible, parallel tasks.

The lesson for engineers is clear: current AI agents are tools, not teammates. They lack the social cognition for organic collaboration. Effective deployment requires architectures that impose order, delegate precisely, and prevent the unproductive loops that naturally emerge when these powerful tools are left to their own devices.

Source: Science News