Cohere Labs' ATE Dataset Reveals Only 2.6% of AI Agent Tools Actually Work

The pitch for AI agents usually goes something like this: hand a bot a job, watch it do the job, count the savings. A new open dataset from Cohere Labs suggests the reality is far messier and far narrower than that framing implies.

The team assembled the Agentic Task Ecosystem (ATE), a corpus of roughly 696,000 published tools across 123,000 MCP servers, the largest open dataset of its kind. Then it ran a strict test: can any given tool actually finish a recognized occupational task on its own, rather than just assist a human doing it? Only 2.6% cleared the bar.

What ATE actually measures

MCP, the Model Context Protocol, is the open standard that lets a language model reach outside its chat window and call external software. One server might expose GitHub, another a calendar, another an internal payments system. Each server ships a bundle of tools with short descriptions of what they do, making them a surprisingly useful public record of what developers believe is ready to hand off to a machine.

Cohere Labs collected 696,291 tools from 123,069 public MCP server listings across seven directories in May 2026, deduplicated the servers, and matched each tool against task statements in the U.S. Department of Labor's O*NET occupational database. For each tool, they identified the closest O*NET task statement and asked a language model whether the tool executes that task outright, filtering out anything that only informs a user or handles a single step in a larger human-coordinated process.

Why 2.6% is probably a floor

Companies build MCP servers against their own internal systems and never publish them. That missing tier skews toward exactly the kind of back-office processes occupational databases describe. Because the count can only miss tools rather than overcount them, 2.6% reads as a lower bound.

The coverage gap across the labor market is equally striking. Of 923 occupations, 419 show no agentic tool activity whatsoever. The supply side of the agent economy is, so far, deeply concentrated.

What the other 98% of tools actually do

Grouping unmatched tools by similarity produced 1,136 categories that fall into three buckets:

• Existing work at the wrong grain. 693 categories are subatomic: recognizable work broken into smaller units than occupational databases use, such as dozens of tools that each handle one sliver of "maintain configuration control."
• Composite workflows. Another 411 categories bundle several O*NET tasks into workflows no single task statement names, like end-to-end meeting-transcript management.
• Infrastructure for agents themselves. A substantial share of the corpus consists of tools that exist so agents can operate: registering agents, discovering other agents, managing sessions, handling identity between them.

Genuinely new categories of human work are rare. Only 35 categories, about 3%, describe work with no plausible counterpart in the occupational record, and most of them exist because agents do: selecting synthetic voices, developing AI personas, assessing whether an agent can be trusted.

Theory versus what got built

Cohere compared its supply-side map against the widely cited Eloundou exposure scores, which rate how much of a task an LLM could plausibly speed up. Across 178 occupations, theoretical exposure and realized MCP coverage correlate at 0.54, respectable for two measures built from completely different data.

A subtler question reveals a sharper divergence: does exposure predict which parts of a job get automated? It doesn't. How much tooling an occupation has says nothing about whether that tooling touches its routine edges or its specialized core. Worker preferences from Stanford's WORKBank survey, measuring what people would actually like automated, predict essentially nothing about what gets built.

Where tools land inside a job

For anyone trying to gauge how AI will reshape a particular field, this is the finding that matters most. The authors borrow an expertise framework from Autor and Thompson to distinguish two patterns of automation:

1. Expertise-lowering: tools handle the specialized core, leaving humans the routine remainder.
2. Expertise-raising: tools handle the routine edges, leaving humans the specialized core.

The pattern varies sharply by domain. In healthcare and computing occupations, available tools reach toward the specialized end of the job, leaving humans the more routine remainder. In production and legal occupations, tools stay at the routine edges, raising the barrier to entry for newcomers. Clinical Data Managers and Biostatisticians, with 143 and 82 matched tools respectively, represent the most extreme cases of specialized-core automation.

The unifying thread is medium, not difficulty. Specialized work is hard to automate where it is physical or interpersonal, and tractable where it is already conducted through software. If your expert judgment is exercised through a keyboard, the agent economy is already coming for it. If it lives in hands, faces, or physical space, current tooling leaves it largely untouched.

Implications for anyone building with agents

Public MCP registries skew heavily toward developer tools, and the long tail of occupations is essentially untouched, which means the space for genuinely useful vertical agents remains wide open. Coverage of graphic designers, for example, is dominated by a single generic task, "Use computer software to generate new images," which dozens of general-purpose tools can technically claim.

A second-order concern deserves attention. Entry-level workers typically learn through the routine tasks that experienced colleagues hand off to them. Hiring of young workers in highly AI-exposed occupations has already fallen behind their peers and the gap has widened over the past year. When routine work is the first to be automated in fields like law and production, the pipeline that produces the next generation of experts grows fragile. The current tool ecosystem does nothing to address that, and one explicit goal of the ATE dataset is to make the problem measurable enough to act on.