Oct 08, 2026
The Agentic Task Ecosystem: A Supply-Side Record of Agentic Automation
The Agentic Task Ecosystem (ATE) is a tagged dataset of 694,411 public MCP tools that reveals developers' tool usage patterns across 178 occupations, with only 2.6% cleanly mapping to O*NET tasks and coverage varying by job specialization and digital information centrality.
Authors
Zanele Munyikwa, Campbell Lund, Thomas Euyang, Aidan Peppin, and Marziah Fadaee
Abstract
Large language models increasingly use external tools to act on digital systems, but there is little systematic evidence about which kinds of work developers are making available to them. We present the Agentic Task Ecosystem (ATE), a tagged, open dataset of 694,411 deduplicated tools from 123,969 public MCP server listings. As a supply-side measure, ATE is an early signal, both preceding and complementary to evidence on AI exposure and usage data. To build ATE, we combine top-down matching to ONET tasks with a bottom-up taxonomy of the full corpus to study both human and machine work for machines. Under our strict matching standard, only 2.6% of tools map cleanly onto a complete ONET task. The bottom-up taxonomy shows what this match rate leaves out. Some categories represent task components no O*NET predecessor, while others span or bundle several recorded tasks. A small set describes work involving the management of agents. Throughout, the units that describe human work (task statements) and machine work (tool descriptions) do not align: machine-performable units arrive both finer and coarser than recorded tasks. Task-level exposure measures therefore bound automation activity rather than count it.Across 178 occupations, we find that matched tasks fall near the midpoint of an occupation's expertise distribution on average, and the amount of tool coverage falls. However, this average conceals differences across occupational groups. Tools reach relatively specialized tasks in computing and legal occupations and relatively routine tasks in production and healthcare occupations. Further analysis suggests that tools reach further into specialized work when working with digital information is central to an occupation's expertise.ATE is bounded by O*NET's incomplete representation of recent work and by our focus on public tools. Future research should examine expertise beyond direct matches, how tools chain into longer workflows, and how these patterns extend into bespoke internal workflows.
Related works
