刚刚，Google DeepMind团队Nature发文：重新定义AI Agent治理

TL;DR - A Google DeepMind Nature perspective proposes “Agentic Profiles,” a framework for governing AI agents by autonomy, efficacy, goal complexity, and generality. It matters because agents with similar model capabilities can pose very different risks depending on their tools, permissions, and deployment environments.

• Replaces binary “agent/non-agent” labels with four multidimensional, graded attributes.
• Demonstrates the framework using AlphaGo, ChatGPT-3.5, tool-enabled Claude 3.5 Sonnet, and Waymo.
• Links each dimension to governance needs such as action logs, scalable oversight, kill switches, interpretability, and cross-sector coordination.
• The framework remains conceptual; quantitative metrics, grading standards, and accountability mechanisms are unresolved.