When Millions of AI Agents Meet
In this DeepMind podcast episode, host Hannah Fry speaks with senior staff research scientist Nenad Tomašev about the transition from large language models to AI agents, and what happens when agents begin transacting, negotiating, and delegating to one another at scale. Tomašev explains that agents differ from LLMs by observing world state, planning multi-step actions, and wielding tools, but stresses that current systems are far from AGI and remain best at combinatorial recombination of existing human knowledge rather than truly novel discovery. The conversation dives into the risks of multi-agent deployment, including prompt injection, hidden tokens in web pages, dynamic cloaking of sites for agents versus humans, and the broader concept of "agentic traps." Tomašev argues for defense in depth: layered mitigations spanning web content attestation, model-side safeguards, permission scoping, and meaningful human-in-the-loop review to combat automation bias. A major thread concerns an emerging agentic economy in which personal assistants negotiate on users' behalf, requiring careful auction design to avoid correlated failures. Tomašev warns of cognitive monoculture and collusion risks that could mirror or exceed flash-crash dynamics, noting that financial markets offer prior lessons. He also challenges the AGI framing: rather than one monolithic general intelligence, the likely future is a distributed society of specialist agents orchestrated by a general connective layer, mirroring how human expertise actually works. This raises new alignment challenges, since safety must be addressed at both individual agent and multi-agent system levels, potentially through economic incentive design.
Original sources
YouTube Google DeepMind
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.