Cohere Publishes Agent-Tool Dataset, Finds 419 Occupations Have No Coverage

The public market for AI-agent tools is vast, but its direct footprint on recognized occupational work is still narrow. Cohere Labs’ new Agentic Task Ecosystem dataset finds that 419 of 923 O*NET occupations have no represented agentic-tool activity, while only 2.6% of the nearly 700,000 catalogued tools meet its strict test for completing a recorded work task end to end.

That result comes from a new public dataset, not an employment measure. Cohere collected 696,291 tools listed on 123,069 public Model Context Protocol, or MCP, servers across seven directories in May. MCP is an open standard that connects AI systems to external software and data, allowing an agent to call functions such as querying a database, editing a file, or scheduling a meeting.

A deliberately demanding definition of automation

Cohere matched each tool description to the closest task statement in the U.S. occupational database O*NET, then used a language model to assess whether the tool executes that task. The protocol excludes tools that merely inform a person or perform one step while a person still coordinates the wider process. Under that definition, 2.6% of tools matched a full O*NET task, covering 1,380 task statements—about 15% of the software-performable work O*NET records.

The unmatched tools are not simply irrelevant

The low match rate does not mean the rest of the catalog is useless. Many tools describe existing work at a different level of detail than O*NET: some break a recognized task into smaller steps, while others bundle several tasks into a workflow. The remaining groups include infrastructure for agents themselves and a smaller category of work that appears to exist because agents do.

How Cohere classifies work outside a full O*NET task match

• 693 categories were “subatomic”: recognizable work divided into units smaller than an occupational task statement.
• 411 categories were composite workflows spanning multiple recorded tasks rather than one task statement.
• Only 35 categories were classified as having no plausible occupational counterpart; most concerned managing or evaluating agents.

Exposure can predict volume without revealing the target

Across 178 occupations, Cohere found a 0.54 correlation between theoretical AI-exposure scores and the number of MCP tools available. But exposure did not predict whether tools landed on routine tasks or specialized ones within a job. That distinction matters: tools reached toward specialized work in healthcare and computing, while legal, production, and sales tools clustered nearer routine work.

The pattern also follows what builders think is technically feasible, rather than what workers say they want automated. For 65 occupations covered by WORKBank data, available tools tracked expert capability judgments but showed no relationship to workers’ automation preferences.

Availability is not deployment

ATE measures the public supply of tools, not their use, reliability, or ability to complete long-horizon workplace assignments. It also excludes bespoke internal MCP servers, which may be concentrated in back-office processes. Cohere argues that omission makes the 2.6% figure a floor for publicly observable tooling, but the dataset cannot establish how much of the listed automation has reached real workplaces.