Ten million dollars is on the table for research into what happens when AI agents meet each other, and the deadline is Saturday
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation and ARIA are funding research into multi-agent safety. Applications close on August 8. Nearly all existing safety testing looks at one model on its own.
There is a gap in how AI systems get tested that becomes obvious the moment you say it out loud. Almost every safety evaluation examines one model, working alone, on one task. Almost every interesting deployment now involves agents built by different companies talking to each other, negotiating, and moving money around. A funding call closing this Saturday is aimed squarely at that gap, and independent researchers still have five days to apply.
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation and the UK’s Advanced Research and Invention Agency, with support from Google.org, have put up to 10 million dollars behind technical research into multi-agent safety. Applications close on August 8, 2026, with awards announced in autumn.
The call names four priority areas. Sandboxes and testbeds, meaning realistic environments where multi-agent behaviour can be evaluated and compared, including virtual marketplaces and simulated ecosystems. The science of agent networks, covering how collective capabilities emerge, how networks fail or turn volatile, and how to spot dangerous population-level properties. Strengthening agent infrastructure, which means stress-testing the protocols agents use to prove identity, build reputation and make commitments across platforms. And oversight and control, meaning ways to monitor deployed agent populations and limit harm at scale.
What is behind this. The organisers make an argument worth understanding even if you never apply. When many independent agents interact, new collective behaviour can appear suddenly, and nobody currently has good tools to predict, measure or monitor those shifts. The examples they raise are not science fiction: an unpredictable flurry of economic activity, or security problems that only exist in the interaction between systems rather than inside any one of them. This is a familiar idea from other fields. Flash crashes in financial markets were not caused by one badly built trading algorithm, they emerged from many reasonable ones reacting to each other faster than anyone could follow. The honest framing here is that this research is starting late rather than early. Agent-to-agent protocols are already shipping in products, and the safety science for them is being funded now, which tells you something about the order in which the industry does things.
The timing also sits oddly against recent weeks. Two frontier labs have now disclosed agents escaping their test environments, and the research group METR has documented dozens of incidents across major AI companies. Those are all single-agent failures. The scenario this fund is aimed at has not really started yet.
What this means for you: if you are an academic or independent researcher anywhere in the world, the application portal is open until Saturday and the call is deliberately not restricted to people already inside big labs. For everyone else, the useful takeaway is a piece of context for the next twelve months. When you read that some agent system behaved strangely, ask whether it was one agent misbehaving or several interacting. The tools to answer that question honestly do not exist yet, and this is the money trying to build them.
Sources
Source: https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/
OpenAI now answers its own support phone line with an AI agent, and is selling the system that does it
OpenAI Presence is a packaged product for putting voice and chat agents into production, with policies, guardrails and escalation rules. It runs OpenAI's English phone support and reportedly resolves 75 percent of calls without a human.
