AI Agents Cross the Threshold: How 65% of Tools Now Take Real-World Actions

A 16-month analysis of 177,000 MCP tools reveals a 2.4x surge in action-capable agents, jumping from 27% to 65%—a shift from passive data access to active intervention where agents now modify files, execute financial transactions, and deploy code. This fundamental transformation exposes the limits of traditional model-level safety; the genuine risk surface has shifted from what AI systems say to what they actually do. Regulators and developers must abandon model-output monitoring in favor of tool-layer authorization and execution oversight.