What happened
Cohere Labs released the Agentic Task Ecosystem, a dataset built from 696,291 tools listed across 123,069 public Model Context Protocol servers. The accompanying analysis examines which occupational tasks developers are building AI agents to perform and finds that agentic tooling is concentrated, unevenly distributed and not yet evidence of widespread adoption.
Cohere Labs says it collected 696,291 tools from 123,069 public MCP server listings across seven directories in May 2026 and deduplicated servers listed in multiple places. The company describes the resulting Agentic Task Ecosystem as the largest open dataset of its kind, but the source does not provide a download location, license or access conditions. Pricing is not documented.
Using a language model to match tools with O*NET task statements, the researchers applied a strict test: a tool had to execute a recognized occupational task from end to end, rather than provide information or complete only one step that a person still coordinates. Under that test, 2.6% of tools qualified. Cohere reports that 419 of 923 occupations had no represented agentic tool activity.
The analysis groups most unmatched tools into smaller pieces of existing work, combinations of multiple recorded tasks, or infrastructure needed to operate agents. Cohere identifies 35 categories—about 3% of the categories examined—as apparently new work with no plausible occupational counterpart, mostly involving agent management such as selecting synthetic voices, switching AI personas and assessing whether an agent can be trusted.
Cohere reports that realized tooling correlates with theoretical AI exposure across 178 occupations, but exposure does not predict whether tools reach routine tasks or specialized work. The source says tools extend toward specialized work in healthcare and computing, while they cluster toward routine edges in legal, production and sales occupations. These are findings reported by the authors; the source does not establish downstream economic effects.
Why it matters
The research offers an early supply-side view of AI automation: what developers have packaged for agents to do, rather than what companies have deployed or workers currently use. Cohere’s analysis suggests that overall exposure measures do not reveal which parts of jobs are affected. That distinction could matter for wages, hiring, training and how workers gain expertise, especially when AI reaches specialized software-based work.
The dataset measures supply, not adoption, reliability or economic impact. A public tool shows that a developer considered a task concrete enough to package for an agent, but it does not show that a company uses the tool, that it works reliably in production or that workers have been displaced.
The source argues that the location of automation inside a job may be more consequential than the total share of work exposed. Removing routine work could leave employees with more specialized responsibilities, while automating specialized work could reduce the expertise required for the remaining job. Either pattern could affect wages and employment differently.
Cohere also reports that expert judgments of technical feasibility predicted which occupations received tools, while workers’ stated preferences about what they wanted automated did not. That finding raises a practical governance question: whether AI development will follow what workers and affected communities value, or primarily what developers judge technically tractable.
The research has a significant visibility limitation. It covers public directories and omits bespoke internal MCP servers, which Cohere says may be concentrated in back-office processes. The reported 2.6% share is therefore presented by the authors as a floor, not a complete measure of automation.
What to watch next
Researchers, employers and policymakers will need to compare the dataset with private enterprise deployments, actual usage, employment and wage data. The most important unresolved issue is whether tools that remove routine entry-level tasks also reduce opportunities for workers to develop expertise.
The dataset’s usefulness will depend on whether other researchers can access and inspect it, reproduce the classifications and test how quickly the public tool landscape changes. The source does not state the release’s license, documentation beyond the article or maintenance schedule.
Future evidence should connect public tooling with private deployments, worker use and measurable labor-market outcomes. Without those links, the ATE findings cannot show whether an occupation is economically threatened or whether a tool improves productivity.
The source highlights a particular risk around entry-level work: if routine tasks are automated, workers may have fewer opportunities to learn through those tasks before taking on specialized responsibilities. Whether organizations redesign training and career paths is an open question.
The analysis also warrants scrutiny because the occupational matches and category groupings rely partly on language-model judgments. Cohere reports a blind human check of 120 categories that pointed in the same direction, but the source does not provide enough detail here to assess the full validation process or error rate.