Summary: Cohere Labs’ Agentic Task Ecosystem dataset maps 696,291 tools from 123,069 public MCP servers collected across seven directories in May 2026. On a strict end-to-end test, only 2.6% of tools could execute an O*NET task completely, coverage was absent for 419 of 923 occupations, and technical feasibility predicted tool building better than worker preferences.
Every MCP server you publish amounts to a quiet wager: that one particular piece of work can be handed to a machine. Cohere Labs has now gathered nearly 700,000 of those wagers into a new dataset, the Agentic Task Ecosystem (ATE), and what emerges from it is far more layered than the familiar story of agents coming for our jobs.
The dataset pulls together 696,291 tools from 123,069 public MCP servers across seven directories, all collected in May 2026. No open dataset of its kind is larger—though the more revealing question is where automation is actually heading.
Only 2.6% of tools genuinely do the work
Cohere’s team ran every tool against the task statements in O*NET, with a language model judging whether it can execute a task end-to-end rather than merely inform the person doing it. On that strict test, just 2.6% of tools clear the bar. About one in forty.
Read that as a measure of the ecosystem’s maturity, not a verdict on it: it says what developers consider ripe for full automation. The other 98% sorts into three groups:
- Subatomic tools (693 categories): existing work broken into units finer than occupational databases record—a task like “maintain configuration control,” for instance, may end up spread across dozens of narrow tools.
- Composite tools (411 categories): the reverse move, stitching several recorded tasks into one workflow—meeting-transcript management is a case in point, since it collects, organizes and analyzes material that O*NET lists separately.
- Infrastructure and new work: tools whose whole purpose is keeping agents running—registering them, discovering them, managing sessions—alongside a thin slice (35 categories, about 3%) of genuinely new work, most of it about managing agents: selecting synthetic voices, switching AI personas, assessing agent trustworthiness.
Tools get built where building is possible
Cohere set theoretical exposure scores (from Eloundou et al.) against actual tool coverage across 178 occupations. They line up at a correlation of 0.54—strong, given how different the underlying evidence is. Exposure scores do forecast how much tooling an occupation attracts.
What they cannot forecast is which parts of a job get automated. Measure where matched tasks sit inside an occupation’s mix—routine edges against specialized core—and the correlation with exposure is indistinguishable from zero. Heavy tooling says nothing about pressure on the core.
Worker preferences fare worse still. For 65 occupations, WORKBank asked what workers would like automated; what gets built instead tracks expert judgments of technical feasibility. Builders are not snubbing workers—they are constrained by what is possible.
Where the coverage stops
Of the 923 U.S. occupations, 419 have no agentic tool activity at all. Ask which jobs agents are coming for, and the supply-side reply is blunt: for a great many of them, no one is building anything yet.
Coverage clusters in software-heavy occupations. Graphic designers, for example, have tools matching 11 of their 15 software-performable tasks—yet over a thousand tools hang off a single broad statement, “Use computer software to generate new images.” That is not deep automation; it is a shallow pool around one generic capability.
In healthcare and computing, tools target specialized tasks and leave the routine remainder to humans. In legal, production and sales, tools stay out at the routine edges and the specialized core stays with humans. Specialized work resists automation where it is physical or interpersonal, and gives way where it is already conducted through software.
What builders should take from this
ATE is a supply-side signal: what has been made available, not what gets used. That is both where its value lies and where it stops. Internal servers companies build for their own back-office processes never surface in public directories, so the 2.6% is best read as a floor, not the full picture.
The practical lesson: the ecosystem is young, and the tools out there lean toward what is technically feasible, not what workers want. You can differentiate by targeting occupations with zero coverage—or by taking on the routine tasks workers want off their plates, which the market has so far ignored.
As a map of where automation is heading, the ATE dataset is genuinely useful. Like any map, though, it shows the ground already covered, not the territory that remains.