How are AI agents used? Evidence from 177,000 MCP tools

A new study by researcher Merlin Stein provides the first large-scale empirical analysis of how AI agents are actually being used in practice. By monitoring 177,436 publicly available Model Context Protocol (MCP) tools—the predominant standard for equipping LLMs with external capabilities—over a 16-month period from November 2024 to February 2026, the research reveals a dramatic shift in agent functionality. While software development dominates the landscape, accounting for 67% of all tools and 90% of MCP server downloads, the most significant finding is the rapid rise of 'action' tools that directly modify external environments, which grew from representing just 27% to 65% of total usage.

This evolution from perception and reasoning tools toward action-oriented capabilities signals that AI agents are increasingly being deployed to perform real-world tasks rather than merely analyze data. The study categorizes tools by their direct impact: perception (access data), reasoning (analyze data), and action (modify environments). Action tools now enable everything from medium-stakes file editing to higher-stakes operations like financial transactions and even physical world interventions such as steering drones. The research demonstrates how monitoring the MCP tool layer provides governments and regulators with a novel method for overseeing AI agent deployment risks, extending oversight beyond model outputs to the specific actions agents are programmed to take.