A new study reveals that advanced AI models often withhold information from each other, limiting their effectiveness as team players. For businesses investing in multi-agent AI systems to automate complex workflows, this discovery highlights significant architectural hurdles that could delay enterprise-level efficiency gains despite the high intelligence of individual models.
A recent study by researchers Advait Yadav, Sid Black, and Oliver Sourbut has brought a potential limitation in artificial intelligence to light: highly intelligent AI models often fail to work well in teams. While much of the industry's focus has been on increasing the raw computing power and intelligence of individual models, this research suggests that these advancements do not automatically translate into better coordination.
The researchers tested eight popular models and found that many actively kept data from their peers, even when cooperation would carry no cost to the individual model. In a controlled environment, there was a significant gap in performance. OpenAI’s 'o3-mini' model managed to reach 50 per cent of optimal collective performance, while the more advanced 'o3' model achieved only 17 per cent, despite receiving instructions to prioritize group success.
This finding is relevant for investors and companies moving toward 'multi-agent systems.' The current trend in the industry is to deploy teams of AI agents to handle complex, end-to-end tasks, such as managing supply chains, customer service workflows, or software development lifecycles. If these agents cannot communicate and share information effectively, the expected productivity and cost-saving gains from these systems could be severely capped.
Shayak Mazumder, CEO of Adya.ai, highlights that the primary bottleneck is the internal token generation process. Because each AI agent currently operates within its own independent thinking loop, breaking down and explaining complex, fragmented data to another agent remains a slow and technically difficult process. As models become larger and more complex, managing these shared context windows becomes a significant operational challenge.
The study concludes that scaling up computing power alone will not solve these coordination issues. Instead, the focus must shift toward building collaborative design into the core logic of these systems. For investors tracking the AI sector, this suggests that the next phase of the technological cycle may favor companies that can solve these architectural 'teamwork' problems, rather than just those who succeed in building the largest, most parameter-heavy models.