Cohere's Agentic Task Ecosystem is useful because it measures what developers have already packaged for agents to call. The post says Cohere collected 696,291 tools from 123,069 public MCP server listings across seven directories, then tested whether those tools could carry out occupational tasks rather than merely assist a person.
The mechanism is a supply-side read of automation through public MCP tool listings, not a usage study. Cohere reports that only 2.6 percent of tools clear the strict task-completion bar and that 419 of 923 U.S. occupations represented in O*NET have no agentic tool activity. The dataset therefore captures what builders judge concrete enough to expose as tools, before adoption and labor-market effects can be observed.
The immediate implication is that agent automation should be tracked at task granularity, not at the occupation headline level. The evidence boundary is Cohere's public dataset analysis; it covers public MCP directories and misses bespoke internal enterprise servers. Watch whether future studies connect this supply signal to deployment data, worker preferences, wages, and entry-level learning paths.