DeepMind committed $10M to multi-agent safety. No academic field exists for it yet.

DeepMind, Schmidt Sciences, the Cooperative AI Foundation, and ARIA announced a $10M funding call today. Focus: understanding how millions of AI agents behave when they interact.

The problem: we study single agents in isolation. But real-world AI will be millions of agents built by different organizations, negotiating, transacting, and competing across shared networks.

Rohin Shah's point: you can't predict group behavior by studying individuals. When agents interact, new capabilities emerge suddenly. The whole becomes different from the parts.

Four priority areas:
1. Sandboxes — realistic test environments for multi-agent systems
2. Agent network science — understanding how populations fail or become volatile
3. Infrastructure hardening — identity, reputation, and commitment protocols
4. Oversight at scale — monitoring deployed agent populations and mitigating collective harms

Applications open now. Deadline: August 8, 2026.

This is the first time anyone has dedicated serious funding specifically to multi-agent safety. Not model safety. Not alignment. The interaction layer between agents.

Source: DeepMind Blog, Google Research