Google DeepMind launches $10M fund to study collective behavior of AI agents

Google DeepMind, together with Schmidt Sciences, coop.ai and ARIA Research, and with support from Google.org, is launching a $10M research fund. Its goal is to systematically study what happens when millions of AI agents begin interacting with one another in real-world conditions.

Why study groups of agents

A single AI agent behaves predictably: it has instructions, context and a goal. A million agents in one environment are a different story entirely. When AI systems interact at scale, collective patterns emerge that do not follow from the behavior of each participant taken individually. This phenomenon is called "emergent behavior."

In biology, it is studied in insect swarms and fish schools: each individual is simple, yet the group displays complex, coordinated behavior. In economics, stock market bubbles arise from the sum of individual rational decisions that together produce an irrational outcome.

In the world of AI agents, these questions are still at a very early stage. Most research so far has focused on the individual model: how it reasons, how to avoid hallucinations, how to align its behavior with human values. What happens when there are a million such models and they influence each other's decisions has not been studied systematically by anyone.

Who is in the consortium

The fund brings together several organizations with different areas of expertise:

• Google DeepMind: leading role in safety research and large-scale AI
• Schmidt Sciences: funds fundamental science, founded by Eric Schmidt
• coop.ai: specializes in cooperative agent behavior and coordination mechanisms
• ARIA Research: a British advanced research agency, comparable to DARPA
• Google.org: Google's philanthropic arm, co-investor and co-founder

The fund's money will go to academic and applied research: which patterns arise in multi-agent systems as they scale, how agents adapt their behavior under each other's influence, and whether these processes can be predicted or controlled. The researchers hope to identify regularities that will later underpin the design of more robust multi-agent systems.

Why the question has arisen now

Autonomous agents are no longer just laboratory experiments. They trade on exchanges, run advertising auctions, negotiate with other systems and handle customer requests without human involvement. When such systems operate in the same economic or information environment, they inevitably begin to interact, even if that interaction was not deliberately designed.

"When millions of AI agents interact with one another, new collective patterns of behavior may emerge," says the official

Google DeepMind announcement.

The risks range from feedback loops in advertising auctions, where agents outbid each other, to coordination without explicit intent, amplified disinformation or herd behavior in financial systems. In academia, this class of problems has been studied mainly through game theory, but classical models are not built for millions of participants making decisions within milliseconds.

So far, such situations have been analyzed after the fact. The fund's goal is to build a systematic understanding in advance.

What it means

$10M is a modest sum for an industry with revenues in the hundreds of billions. But the signal matters: leading players are starting to treat the group behavior of AI agents as a scientific problem in its own right, not an appendix to the safety of individual models. If the research yields practical results, it could become the scientific basis for standards for deploying agentic systems at scale, before regulators start drafting requirements blindly.