Google DeepMind has teamed up with Schmidt Sciences, the Cooperative AI Foundation, the UK’s Advanced Research and Invention Agency, and Google.org to launch a new funding call aimed at studying the safety risks of multi-agent AI systems.
The initiative offers up to $10 million for researchers around the world and is intended to accelerate work on how large numbers of autonomous AI agents behave when they interact with one another. The companies and organizations behind the program say that as AI systems become more widespread, they will increasingly operate in environments where many agents, built by different developers, can communicate, negotiate and transact.
The funding call focuses on research that can help explain and manage the system-level risks that may emerge when these agents act together. According to the announcement, current safety methods are largely designed to evaluate AI models on their own, rather than as part of a broader network. The new program aims to close that gap by supporting work on behaviors that may be difficult to detect when systems are tested individually.
Researchers involved in the effort say the field is moving into a new phase in which collective behavior matters as much as the performance of single models. They argue that interactions among agents can produce sudden shifts in capability or behavior that are not obvious in advance. Potential concerns raised in the announcement include unpredictable economic activity, security risks and other population-level effects that could arise when many systems interact across digital environments.
The call builds on earlier work from Google DeepMind and others on multi-agent systems. The company points to its 2025 research framework for studying these interactions, along with more recent work on so-called AI Agent Traps, which examined vulnerabilities in adversarial settings. The new funding round is designed to move from foundational concepts toward broader empirical research, with an emphasis on practical tools and shared standards.
Applications are being sought from academic teams and independent researchers. The organizers have outlined four main research areas. One is the development of sandboxes and testbeds, including simulated marketplaces, ecosystems and multi-organization workflows that can be used to study agent behavior in realistic settings. Another is the science of agent networks, which would examine how collective capabilities emerge, how groups of agents scale and how networks can become unstable or fail.
A third area is agent infrastructure, including work on identity, reputation and commitment systems that support secure interactions across platforms. The fourth is oversight and control, with a focus on monitoring deployed agent populations and reducing harms when they arise at scale.
The organizations behind the funding call say a diverse research community is needed to build transparent and robust safety standards for the emerging agent ecosystem. They also frame the effort as part of broader work by Schmidt Sciences and ARIA on trustworthy AI and multi-agent coordination.
The deadline for proposals is August 8, 2026. Awardees are expected to be announced in autumn 2026.
The new funding call comes as major AI labs and research funders increasingly turn attention to what happens when AI systems no longer operate in isolation. With agentic tools becoming more capable, the industry is starting to look beyond single-model safety and toward the risks created by systems working together.