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 question is empirical, and the field cannot answer it, because its record of intervention is content-blind: in agentic coding, the best-measured setting, 91.49% of failure resolutions in the largest published analysis required explicit user correction, against 2.99% agent self-correction, yet every record registers the correction as a single undifferentiated event. Frequency says nothing about which interventions capability growth can remove. Existing taxonomies index interventions by timing, authority level, granularity, defect class, or interaction state. The missing axis is epistemic content: whether the input supplies a fact, a frame, a pointer, a decision, or nothing but a demand to commit. From seven months of my own practice with LLM coding agents, documented in structured session retrospectives, I derive nine recurring mechanisms indexed by that content, with a lower-evidence second tier. Three of the nine carry no domain content, moving only commitment, salience, or attention: authority exercised, not knowledge transferred. What a mechanism supplies determines what can absorb it. Mechanisms supplying grounds (evidence bearing on entitlement to conclusions) are absorbable by capability growth. Mechanisms supplying frames (replacements for the hypothesis space itself) yield to a second reasoner, which need not be human, though the fraction whose salience forms in bearing the outcomes absorbs last, if at all. Mechanisms supplying standing require a principal, and standing cannot be self-conferred, so no capability increment absorbs them. Stated formally under incomplete contracting, expected loss from a misset delegation threshold converges to a strictly positive floor equal to the conditional dispersion of the principal's threshold, the compositional prediction follows as a corollary, and the third class is exposed as conditional on non-enumerability that persists under learning. The partition is an oversight result: a stop internalized as a preference transmits no standing, the delegation-side form of the corrigibility problem, and trained deference that rewards wrong corrections degrades the oversight signal itself. In the limit, an agent granted authority, accountability, and diverse priors absorbs everything but proprietorship: administering a success criterion is delegable, originating it is not, and the residue rests with whoever bears the outcomes. One institutional corollary: every external lever of control, liability included, presupposes a party that can be worse off, which derives the current priority of training-side alignment from institutional structure rather than capability risk. Across eleven literatures, twelve of thirteen mechanisms carry established names, and the taxonomy works as a concordance across fields that do not cite one another. What no surveyed literature claims is the direction: every adjacent field builds machinery that makes humans think harder, and none studies the human as the forcing function applied to a machine's live reasoning. The corpus is one practitioner's records, a discovery instrument rather than evidence. Validation is specified as re-coding public interaction corpora, where cause-for-cause correspondence with an independently derived 20,574-session failure taxonomy already provides preliminary convergence.