Automation’s Early Footprint

Hundreds of thousands of tools have been published to let AI systems take action on our behalf: write code, query databases, update records, schedule meetings, move information between applications. But what does this rapidly growing layer of agentic AI infrastructure actually add up to? What work are we building AI to automate?

Most of what we know about AI and work comes from one of two vantage points. Some studies estimate what AI could theoretically do, calculating exposure based on the percent of occupational tasks an AI system could plausibly handle. Others look at what people ask AI to do, analyzing millions of conversations to see which tasks users bring to a chatbot [ AI Observatory, 2026 ; Handa et al., 2025 ; Iscenko et al., 2026 ; OpenAI, 2026 ; Tomlinson et al., 2025 ]. Both represent distinct pieces in the larger puzzle of how AI is impacting work, however, they do not represent the full picture on their own. At Cohere Labs, we are committed to helping assemble this puzzle, and our research presents new evidence from a third, complementary vantage: the agentic-task ecosystem .

We collected nearly 700k tools from 123k public MCP servers to create the Agentic Task Ecosystem (ATE) dataset and ask, what tools have developers built to enable AI agents to fully automate certain work tasks? What we found complicates the idea that agents are simply swallowing occupations one task at a time.

1) Most of it is existing work, seen at the wrong grain. We found that 693 of the categories are what we call subatomic : recognizable work broken into smaller units than occupational databases use. Where a task statement says "maintain configuration control," ATE holds dozens of tools for individual components of that task. Another 411 categories of MCPs run the other way, spanning several recorded tasks or bundling them into a workflow that no single task statement names: meeting-transcript management, for instance, collects, organizes, and analyzes transcripts across what O*NET records as separate tasks. Neither kind falls outside the world of recognized work, the subatomic tools simply do not complete the whole process of what a task statement describes, and the composite ones stitch together several tasks.

In the past two years, AI systems have moved from generating text to taking autonomous actions. A model that once only answered inside a chat window can now query a database, edit a file, open a pull request, schedule a meeting, or run a sequence of these tasks in order.

The Model Context Protocol (MCP) is an open standard for connecting AI systems to external software and data. One server might let a model work with GitHub, another might give it access to a calendar, a spreadsheet, a payment system, or a company's internal documents. In practice, a developer writes an MCP server that consists of a set of tools, each with a short description of what it does.

Economists studying automation usually weigh three variables against each other: (1) what work humans do, (2) what work gets handed to machines, and (3) what genuinely new work appears for both. Automation displaces workers from tasks machines can perform whereas genuinely new tasks that emerge, where humans hold the advantage, reinstate them. Whichever effect is larger determines whether a technology raises or lowers demand for human work overall.