Summary
The Google DeepMind podcast welcomes Nanad Tomashev, senior staff research scientist at Google DeepMind, for a wide-ranging conversation on what happens when AI agents stop working in isolation and begin transacting, delegating, and negotiating with each other at scale. The episode covers the conceptual foundations of agents — how they differ from standalone language models by observing world state and taking actions — before moving into the harder questions that emerge in multi-agent environments.
Tomashev distinguishes between current multi-agent systems, which mostly parallelize work by chunking tasks into independent subtasks, and a more sophisticated delegation model where orchestrating agents must assess the reliability and capabilities of sub-agents, handle failures gracefully, and verify outputs that may not have clean ground truth. He raises the concept of reward hacking in this context: an agent may technically satisfy a request while violating its spirit, which is far more dangerous in agentic chains than in single-turn interactions.
The discussion also addresses the economics of a world with millions of interacting agents — a potential new kind of digital economy — and what trust, certification, and safety mechanisms would need to look like to make it work. Google DeepMind tools mentioned include Gemini Spark and anti-gravity, alongside open-source reference points like OpenClaw. The conversation is grounded in research without being overly technical, making it accessible to anyone trying to understand where agentic AI is headed and what guardrails are still missing.
📺 Source: Google DeepMind · Published June 23, 2026
🏷️ Format: Podcast
