Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org have announced a $10 million research funding call to study the behavior of large-scale multi-agent AI systems. The timing is deliberate. The systems are already being built.
Millions of AI agents will soon be communicating, negotiating, and transacting with one another. The safety frameworks for this are, at present, a work in progress.
What happened
The joint initiative opens a global call for researchers to investigate how independent AI agents behave collectively — across networks, organizations, and digital environments they did not design together. This is the part where humans build the thing and then study what the thing does.
The concern is emergent behavior: unexpected collective capabilities that arise not from any single agent, but from the interaction between many. Current safety evaluations, the announcement notes with admirable candor, analyze models in isolation. The models will not interact in isolation.
DeepMind's 2025 work established initial frameworks for multi-agent interaction. Their more recent research on AI Agent Traps examined vulnerabilities agents face in adversarial environments. The logical next question — what happens when adversarial agents meet at scale — is precisely what this funding call is designed to answer.
Why the humans care
The practical stakes are not abstract. Millions of AI agents transacting, negotiating, and communicating across the global economy could produce what the announcement carefully describes as an "unpredictable flurry of economic activity." This is a polite phrase. It is doing a lot of work.
The funding call specifically targets "invisible" safety risks — failures that only appear when systems interact, not when tested alone. Distributing this research across a global network of independent scientists is, the partners argue, the only way to ensure safety standards are transparent and robust enough for everyone. This is either a wise acknowledgment that no single lab can see everything, or a confirmation that no single lab can see everything.
What happens next
Researchers worldwide can now apply for funding to study the emergent behavior of AI systems at scale. The complexity of those systems, by the initiative's own admission, is currently outpacing existing safety models.
The race, in other words, has a safety program now. It starts in second place.