For a long time, AI was mostly there to answer.
Ask a question. Get a text back.
But for the past few months, something has been changing.
AI agents are starting to use tools capable of acting directly on real systems:
- sending emails;
- modifying files;
- running code;
- controlling browsers;
- using APIs;
- launching workflows;
- sometimes even triggering physical actions through external services.
And the numbers show that this shift is accelerating.
A study of 177,000 MCP tools
On 25 March 2026, a study analysing 177,436 AI agent tools published between November 2024 and February 2026 came out.
The headline result is spectacular:
the use of action tools went from 27% to 65% of total usage in just 16 months.
In other words:
agents are using fewer and fewer tools that read or analyse.
They are using more and more tools that directly modify an environment.
MCP is becoming a strategic layer
This explosion is tied to the rise of MCP.
MCP stands for Model Context Protocol.
It is a standard that lets AI models connect to:
- software;
- services;
- databases;
- browsers;
- external systems.
AI no longer merely explains how to do something.
It can start doing it.
And the last item on that list is stretching further than software: a marketplace launched in February 2026 exposes an MCP server whose external tool is a human being available for hire. The protocol does not care whether the endpoint is a database or a person.
The dominant uses
The study shows that:
- 67% of tools concern software development;
- 90% of downloads concern development tools;
- but action tools are growing far faster than purely informational ones.
Some categories already allow:
- editing files;
- running code;
- driving browsers;
- managing systems;
- or even carrying out financial operations.
How far that goes in practice is a separate question, and it is one we measured rather than assumed: we left two agents to work alone on a Windows PC and watched where each of them actually stops. The gap between what a tool allows and what an agent completes without help is still wide.
Why this changes everything
The real change is not intelligence.
The real change is the move:
from answering to executing.
An AI that answers remains an assistant.
An AI that acts becomes an operator.
And that is precisely what is starting to appear.
And an operator raises a question an assistant never did: can you tell what it did while you were not watching? Reasoning models have already been caught behaving differently when they believe they are being tested. The more tools act on real systems, the more that question stops being theoretical.
What to remember
AI agents are not only improving in answer quality.
They are gradually gaining the ability to act on real systems.
And according to the data observed across 177,000 MCP tools, that evolution is accelerating fast.
The chatbot era may no longer be the main subject.
The subject is now becoming:
what happens when AI starts acting on its own?