An Epistemic-Content Taxonomy of Human Intervention in Agentic Collaboration

When a human intervenes in an agent's work, their input supplies epistemic content; yet existing taxonomies index such moves only by timing, authority, granularity, defect class, or interaction state, not by what the input supplies. Two months of one practitioner's workstation record, 2,205 inputs to LLM coding agents with 1,094 of them gated as interventions, are coded on two axes: the epistemic content of the input, whether grounds, a frame, or an act of standing, and the target the input acts on. Grounds are considerations bearing on entitlement to a conclusion, which live in the world and can be looked up, a frame is a replacement of the hypothesis space the agent is searching, and standing is the authority a move exercises while supplying no domain content of its own, which is what a stop signal, a demand for judgment, or a re-anchoring to purpose transfers. The traffic those schemes leave uncoded is where the corrective load sits. In an inductive analysis of 20,574 agentic coding sessions, 91.49% of misalignment resolutions required explicit user pushback, and the finest-grained published taxonomy of that traffic sorts it into four categories framed as a measure of friction rather than of content, the largest of them defined as corrections in which the user provides missing information or redirects the approach, which is a fact supplied and a search space replaced under one label. Every agreement figure below comes from a panel of language model instances rather than of people, a convention stated once here rather than marked beside each figure. Three boundaries strain, and each strain is measured rather than asserted. A blind re-read of the turns the gate set aside as assent measured what each turn delivered to the agent rather than what it looked like, and 93 percent of the 29 turns read carry non-zero payload, which shows the gate keying on a surface form anticorrelated with the property it screens for. Projected onto the twelve turns not read, that rate admits roughly 38 of the 41 assent-gated turns, and the record states no rule by which the 29 were selected, so the projection rests on the unread turns resembling the read ones rather than on a sampling design. What the gate drops is concentrated rather than scattered. The modal category in that re-read is an authorization releasing an action the agent has named and parked, which is standing where one action was parked, and content and authority at once where several were, the latter sitting in none of the three classes. Sizing that split, the release cell follows turns that left a mean of 1.19 actions open and a median of zero, with 52 percent of it following a turn that enumerated nothing at all. The second boundary is the grounds class's own definition. The class defined as one verifiable fact that collapses a premise is, across its 85 items, 72 percent fact, 20 percent derivation from material already in the conversation, and 8 percent stipulation that the premise does not bind, from three blind coders at 91 percent unanimity, so 28 percent of a grounds-assigned class is not a fact and the assignment overstates what retrieval reaches. The third boundary is the standing class. The strain there is narrower than a first reading suggests. Two raters of different lineage, agreeing at 0.702 on whether a payload is present, find one in 47.8 percent of the standing class's 203 items. A second pass, at 0.734, settled where the payload goes for 58 of those 97 and split them evenly. The derivable half is a conclusion already available in the conversation and not reached, which is what the zero-content claim asserts about these mechanisms. The retrievable half is what the agent did not have: 29 items, 14 percent of the class, or about a quarter if the 39 that no uptake could place divide the same way. A third of the class stays unresolved. The retrievable half also concentrates. Of the demand items whose destination could be settled, the judgment demand is 62 percent derivable of 32 and the warrant demand 76 percent retrievable of 21. The strain is therefore the warrant demand's rather than the whole class's. All three therefore strain by measured amounts rather than in principle. The second and third are one reading run in both directions, putting 28 percent of a fact-defined class outside fact and 14 percent of a payload-free class inside payload. The scheme those three belong to descends from an inventory derived over seven months of the same practice: nine recurring intervention mechanisms with a lower-evidence second tier of four, thirteen in all. Twelve of the thirteen already carry established names in the literatures surveyed, and the recommendation is to adopt those names rather than coin new ones, with every novelty claim scoped to a survey limited by what was reachable and searched on this paper's own terms. Three of the nine are typed as injecting zero domain content, a class-level description the per-item coding only partly corroborates, since that coding marks 5.4 percent of the record pure-authority against the 9.0 percent those three mechanisms carry. The coding uses 28 classes, of which the thirteen are a minority. They cover 407 of the 1,094 interventions, while 555 fall in classes descended from the residue a single-pass coding could not place, and 132 in classes that predate the re-code. Composition over those classes is reported on the two axes separately, as ranges spanning seven independent passes, because reporting them together resolves content for only the third of the record targeted at live reasoning and reports location for the rest. On the target axis the deliverable takes 42 to 59 percent, live reasoning 19 to 42, the rule store 4 to 17, the channel 3 to 6, the authority relation 1 to 2, and the self-model 0 to 2. The authority-relation floor holds over the five main-corpus passes rather than over all seven, since the two extension passes carry no rows in that cell. On the content axis grounds takes 46 to 76 percent, frames 14 to 33, and standing 8 to 22, through a published mapping of the 28 classes onto the three families that is a judgment rather than a measurement, and the 30-point spread on grounds against 17 on the largest target cell states how much less settled the content axis is. The levels move with protocol, and two results do not: the deliverable is the largest target in every pass, and grounds is the largest content family in every pass. The shares are of intervention turns rather than of episodes, since an episode coding finds roughly a quarter to two fifths of interventions continuing an ongoing push on the same issue, and they cover the typed channel alone. Over the two months where both channels are datable, 1,094 interventions were typed while 277 rules were persisted to a version-controlled rule store, so roughly one intervention in five went to a channel no count here observes. Whether raters other than the author can apply it is a separate measurement. One convention governs every agreement figure above and below: rater, coder, judge and reader name one role throughout, and no human second-rater run has been completed, so a coefficient here bounds how consistently a model panel applies a rubric and not how well that rubric survives an independent human reader. Blind pairs drawn from different model tiers reach a Cohen's kappa of 0.77 on whether an input is an intervention at all and 0.70 on its content family once labeling conventions are pinned, against 0.59 with those conventions unpinned and a pilot using two raters from one model family reaching 0.86 as the upper bound. Kappa scores how far two raters agree beyond the agreement chance alone would produce, so 0 is chance and 1 is perfect. The pinned conventions were derived from the first run's disagreements, so the runs are a sequence rather than a design. On sessions held out from the scheme's development the gate replicates at 0.82 while family agreement reads 0.58, both computed across all 93 held-out items. One contrast states what those six figures are figures of. Holding corpus, rater family, and contexts constant and varying only the question, asking raters to forecast whether the next human turn would approve scores at chance, a discrimination gap of minus 14 points, while the same triples agree with one another on 77 percent of items. Asking instead what a turn delivered, once shown it, sorts at kappa 0.965, and that task reproduces in a second model lineage. Reliability here is a property of the question asked rather than of the raters or of the scheme, so an agreement figure carries no information without a validity arm beside it, and each number above is bounded by which question its instrument put. The composition shares inherit none of these figures, since they come from a later, author-adjudicated instrument carrying no reliability figure of its own. One claim above comes from none of those instruments. That no published work indexes on this axis rests on a check rather than on a survey's silence, and the check narrows it. The released data of the nearest published scheme ships a pushback column carrying seven values against the four its paper documents, and two of the three undocumented values fall on this axis, one frames-shaped and one standing-shaped. What is missing is therefore not the distinction but any published measurement of it, since no work states it, reports its distribution, or gives a reliability figure for it. Two further candidates checked since resolve elsewhere: a recent taxonomy of developer-AI collaboration indexes on trigger modality, and CUPS on activity state. None of these measurements reaches past discovery. The corpus is one practitioner's records, and validation runs through re-coding public interaction corpora, which requires none of that record. The nearest such corpus is now released, 5,851 sessions over 205 repositories under ODC-BY with human prompt text and agent replies verbatim after scrubbing