Google DeepMind researcher Nenad Tomasev explains how agents differ from ordinary language models: they observe an environment, plan across multiple steps and take actions through tools. That autonomy can accelerate software and science, but every action still carries a failure rate that makes engaged human oversight essential.
Nenad Tomasev describes a future in which general agents delegate to certified specialists. Reliable coordination requires explicit contracts, reputation, verification and different handling for reversible and irreversible actions. Parallel agents alone do not provide intelligent delegation if work is divided without shared context or accountability.
The safety challenge grows when agents operate on the open web or transact with one another. Prompt injection, dynamic cloaking, collusion and cognitive monoculture can create correlated failures across large systems. Tomasev recommends defense in depth across models, agents, permissions, trusted resources and human controls, alongside economic and policy mechanisms for multi-agent societies.
Watch on YouTube
