Google DeepMind and partners launch $10M multi-agent AI safety research call
Google DeepMind announced a $10 million multi-agent AI safety research initiative on June 11, 2026, joining with Schmidt Sciences, the Cooperative AI Foundation, the UK Advanced Research and Invention Agency (ARIA), and Google.org to fund academic and independent researchers worldwide. The call targets four priority areas: sandboxes and testbeds for multi-agent evaluation, the science of collective capabilities in agent networks, strengthening agent infrastructure, and oversight methods for deployed agent populations. Applications are open through August 8, 2026.
What's new
Partners:
- Google DeepMind
- Schmidt Sciences
- Cooperative AI Foundation
- Advanced Research and Invention Agency (ARIA)
- Google.org
Grant size: Up to $10 million total across awardees
Who can apply: Academic and independent researchers worldwide
Application deadline: August 8, 2026 Award announcement: Autumn 2026
Four funded research areas:
- Sandboxes and testbeds — Environments where multi-agent behavior can be safely evaluated and monitored at scale
- Science of agent networks — Understanding how collective capabilities and emergent behaviors arise when large numbers of agents interact
- Agent infrastructure security — Identity verification, reputation systems, and commitment protocols to make agent-to-agent interactions trustworthy
- Oversight and control — Methods for monitoring and intervening in deployed populations of AI agents
Context
The initiative arrives as AI companies are deploying agents at scale, and as multi-agent architectures — where many AI systems coordinate, negotiate, or compete in shared environments — are moving from research labs into production. Most existing safety evaluation focuses on individual model behavior in isolation.
The collaboration spans academic institutions, private foundations (Schmidt Sciences, Cooperative AI Foundation), government-backed research bodies (ARIA), and philanthropic channels (Google.org), suggesting an attempt to build a multi-sector research base for what DeepMind frames as a structural gap in the field.
The Cooperative AI Foundation's involvement is particularly notable. The organization focuses specifically on cooperation problems in multi-agent settings — whether AI systems can coordinate reliably on shared goals without colluding in ways that harm third parties. That angle is distinct from traditional alignment work focused on single-model behavior.
Why it matters
"Soon, millions of AI agents — built by different organizations — will interact across digital environments, communicating, negotiating and transacting with one another." That framing from the announcement describes an infrastructure challenge as much as a safety one: the AI ecosystem is converging on a world where the relevant unit of analysis isn't a single model but a population of interacting agents operating across many systems simultaneously.
The $10M figure is modest by AI company standards — but the scope of the call is broad enough to fund a meaningful cohort of researchers across multiple institutions. More significantly, the grant creates structured demand for research that doesn't exist in meaningful quantity today: methodologies for evaluating, predicting, and intervening in emergent collective AI behavior.
The August 8 application deadline and autumn award announcement suggests awardees could begin work in late 2026 — timed to a period when agentic deployments are expected to grow substantially as more frontier models add reliable tool use, orchestration, and persistent memory.
Corroborating sources
- Deepmind
https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/
“Soon, millions of AI agents — built by different organizations — will interact across digital environments, communicating, negotiating and transacting with one another.”