When you publish an MCP server, you’re making a quiet bet: that a specific piece of work can be handed to a machine. Cohere Labs has now catalogued nearly 700,000 of those bets in a new dataset called the Agentic Task Ecosystem (ATE), and the picture it paints is more nuanced than the usual “agents are coming for our jobs” narrative.
ATE aggregates 696,291 tools from 123,069 public MCP servers across seven directories, collected in May 2026. It’s the largest open dataset of its kind, but its real value is in what it reveals about the direction of automation.
The 2.6% that actually automates work
Cohere’s team matched each tool against O*NET task statements, asking a language model whether the tool can execute the task end-to-end, not just inform someone doing it. Under that strict test, only 2.6% of tools clear the bar. That’s about one in forty.
This isn’t a failure of the ecosystem. It’s a signal about what developers judge to be ready for full automation. The other 98% fall into three buckets:
- Subatomic tools (693 categories): existing work broken into finer units than occupational databases record. For example, “maintain configuration control” might be split into dozens of specific tools.
- Composite tools (411 categories): workflows that stitch together several recorded tasks, like meeting-transcript management that collects, organizes, and analyzes across what O*NET lists separately.
- Infrastructure and new work: tools that exist so agents can operate—registering, discovering, managing sessions—plus a small slice (35 categories, about 3%) of genuinely new work, mostly about managing agents: selecting synthetic voices, switching AI personas, assessing agent trustworthiness.
What gets built follows what can be built
Cohere compared theoretical exposure scores (from Eloundou et al.) with actual tool coverage across 178 occupations. The correlation is 0.54—strong, given the two measures come from entirely different evidence. Exposure scores do predict how much tooling an occupation receives.
But they don’t predict which parts of a job get automated. The correlation between exposure and where matched tasks sit inside an occupation’s mix—routine edges versus specialized core—is indistinguishable from zero. A heavily tooled occupation isn’t necessarily under pressure at its core.
Worker preferences fare even worse. For 65 occupations, WORKBank surveyed what workers would like automated. What actually gets built tracks expert judgments of technical feasibility, not worker desires. Builders aren’t ignoring workers; they’re simply constrained by what’s possible.
Where the gaps are
Nearly half of U.S. occupations (419 of 923) have no agentic tool activity at all. If you’re wondering which jobs agents are coming for, the supply-side answer is: for many, no one is building anything yet.
Coverage concentrates in software-heavy occupations. Graphic designers, for instance, have tools matching 11 of their 15 software-performable tasks—but over a thousand tools attach to a single broad statement, “Use computer software to generate new images.” That’s not deep automation; it’s a shallow pool around one generic capability.
In healthcare and computing, tools target specialized tasks, leaving humans the routine remainder. In legal, production, and sales, tools stay at the routine edges, leaving the specialized core to humans. Specialized work resists automation when it’s physical or interpersonal, but gives way when it’s already conducted through software.
What this means for product builders
ATE is a supply-side signal: it measures what’s been made available, not what’s used. That’s both its strength and its limitation. Public directories miss the bespoke internal servers companies build for their own back-office processes, so the 2.6% figure is best read as a floor, not the full picture.
For builders, the takeaway is practical: the ecosystem is still early, and the tools that exist are skewed toward what’s technically feasible, not what workers want. If you’re building agentic tools, you have room to differentiate by targeting occupations with zero coverage—or by addressing the routine tasks workers actually want off their plates, which the market has so far ignored.
The ATE dataset is a useful map of where automation is heading. But like any map, it shows the terrain that’s been explored, not the territory that remains.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.