Grounds, Frames, and Standing: Human Interventions in Agentic Work and the Limits of Delegation
When, if ever, is human intervention required for an AI agent to operate successfully? The field cannot answer this, because its record of intervention is content-blind: large-scale studies register each correction as a single undifferentiated event. This paper indexes interventions by epistemic content instead, and argues that what a mechanism supplies determines what can absorb it. Mechanisms supplying grounds are absorbable by capability growth, mechanisms supplying frames yield to any second reasoner with different priors, and mechanisms supplying standing require a principal. Stated formally under incomplete contracting, expected loss from a misset delegation threshold converges to a floor set by the conditional dispersion of the principal's threshold, and the compositional prediction that aggregate exposure measures cannot detect follows as a corollary.
PDF · DOI: 10.5281/zenodo.21563155
Claims are scoped to the literatures surveyed. The discovery corpus is one practitioner's records and is treated as a discovery instrument rather than as evidence, with validation specified against public interaction corpora.