Cohere has assembled roughly 696,000 public AI tools into the Agentic Task Ecosystem, or ATE, a dataset designed to measure what developers are actually trying to automate. It combines seven directories and includes tools from about 123,000 Model Context Protocol servers.
Only 2.6% of the tools passed Cohere’s strict test for automation: carrying out a recognized occupational task rather than merely giving a worker information. Nearly half of the 923 US occupations in the analysis had no agentic tools represented at all. The tools that do exist cluster around narrowly defined work, infrastructure for running agents, and the emerging job of managing agents themselves.
The pattern also varies by profession. In healthcare and computing, tools often target more specialized tasks while leaving routine work to people. In legal, production, and sales roles, automation remains closer to the routine edges. Cohere also found that technical feasibility predicted where tools appeared, while workers’ stated preferences about what they wanted automated did not.
ATE is being released publicly so researchers can track how the mix changes. It measures published tools, not adoption or real-world job displacement, so it is an early indicator of developer activity rather than a direct count of work already automated.