Most studies of AI and work begin with a forecast: Which tasks could a model perform? A new preprint takes a different route by examining public files that people wrote to tell AI agents how to carry out recurring work. Researchers Hyeongjae Lee, Jihyang Cheon, and Lanu Kim collected public skill.md files from the Manus Skills Marketplace. These files describe reusable routines for agents. After filtering the collection for substantive, work-related instructions, the researchers retained 53,515 skills. They compared those instructions with 17,951 task statements covering 748 occupations in the U.S. Department of Labor’s O*NET database.
The result is an Agentic Adoption Index, or AAI. It estimates how closely the public agent instructions resemble the tasks associated with each occupation. The score is comparative. It cannot tell us how many workers use an agent, whether the instructions were executed, or whether the resulting work succeeded.
That boundary is also what makes the study useful. A capability benchmark asks what a model can do. Product availability asks whether a tool exists. Usage telemetry records an interaction. A reusable workflow file captures another layer: someone invested enough effort to specify a task for delegation and share the procedure.
By contrast with a generic list of jobs exposed to AI, the occupational pattern is specific. Management analysts, technical writers, computer programmers, urban and regional planners, and several research and administrative occupations rank high. Jobs dominated by physical handling, close personal care, or highly situated manual work rank low. The researchers also find little overlap with an older automation-risk classification. Using a median split, the Jaccard overlap is 0.05.
Their index correlates more strongly with measures of AI capability and tool availability than with observed Claude usage: Spearman correlations of 0.673, 0.623, and 0.374, respectively. Those figures suggest that configured delegation occupies a distinct stage between technical possibility and established workplace practice. Public workflow artifacts can reveal where early adopters are trying to turn AI into repeatable procedure, even when broad usage data has not caught up.
But one result deserves caution rather than a tidy explanation. Occupations that typically require graduate education show lower AAI values than the model predicts after accounting for tool availability. The paper cannot determine whether this gap comes from tacit knowledge, liability, interpersonal demands, professional norms, or simple reluctance to publish sensitive workflows. It identifies a shortfall while leaving the cause unresolved.
The dataset has other limits. Public marketplace contributors are early adopters, and configuration files record design activity rather than completed work. O*NET describes established occupations and may miss new tasks created by agentic systems. Public sharing also favors work that can be documented without exposing confidential processes.
Verification bottleneck
The useful question is whether a workflow specification became reliable work under accountable supervision.
Before treating a public skill as evidence of adoption, ask:
- Was the workflow executed repeatedly in a real setting?
- What outcome was checked, by whom, and against which standard?
- Which permissions, exceptions, and repair steps remained with a person?
Without those receipts, the file remains a delegation blueprint. With them, it can become evidence about operating practice.
Opportunities
Organizations can use workflow artifacts as a candidate map. Repeatedly specified tasks deserve closer study, especially when the same routine appears across independent teams. The next research layer should connect each candidate to execution frequency, error rates, correction effort, and the authority granted to the agent.
Workers and professional groups have a parallel opportunity: document the judgment that the workflow leaves out. That includes exceptions, tacit cues, responsibility for consequences, and the moments when a task must return to a person. This is where the Human Premium becomes measurable rather than rhetorical.
So the public workflow file is best understood as a trace of attempted delegation, positioned between forecast and proof. Those traces can show where the workplace is being redesigned before formal surveys see it.
Source: Who Delegates to AI? Evidence from 53,000 Agent Configurations (PDF)