Agentic Institutions: Mandates, Capabilities, Constraints, and Escalation for Machine Economic Actors

Delegated cognition, one persistent model dispatching drafting to others, is becoming a common mode of production in agent systems, yet the institutions governing it remain ad hoc. This paper contributes three things: a conditional cost model for delegation, an accounting of context admission, and an instrumented self-study of one institution with its incident record. It does not establish a general institutional design. First, the cost model. For free workers (w = 0) delegation wins whenever the single-attempt acceptance probability p exceeds a small threshold set by the cost of a failed attempt, about 0.02 at the magnitudes of the Table 1 drafting waves (on an ε the instrument does not log and a hub-audited α), and that condition, not the paid-worker horizon, is the result this paper's record supports. A pre-registered wave with admission logged on both arms (§3.8) then found α at 0.165 pooled against the attested 0.137, acceptance 6 of 6 under the rule's enumerated clause, and b/S inside its band, but also found the dispatcher returning each whole draft to the hub as print-back (o/S = 1.007), so the per-context-budget multiplier measured 0.57 against a predicted 2.0 to 3.7 and the free-lane corollary's own proviso failed. The context saving this paper prices exists only when the whole draft does not persist in hub context, and on this instrument it did. For paid workers, with inline characters re-paid on every remaining turn T of a session and delegated characters paid once, delegation has lower expected cost when (w/h) < p(1+T)(1−α) up to overhead terms, where w/h is the worker-to-hub price ratio and α the fraction of worker output the hub admits into its context. That is the rent-or-buy breakeven with a horizon, stated under an assumption the paper names (inline output persists in context for the session) and shows its own hub violates; under full eviction the horizon term disappears and, at zero overhead, the condition reduces to per-query routing. Per unit of context budget, delegation admits up to 1/α artifacts per inline artifact, approached only as the brief and print-back terms vanish (§3.4; 0.57 on the one wave that logged it). Every p and α taken from the run is an illustrative estimate, p hub-attested on 13 briefs and α hub-audited on the 12 woven dispatches, not a measured rate for any worker. Second, it conjectures that as worker cognition approaches zero marginal cost, the binding constraint migrates to hub verification, and it argues that a four-layer institutional anatomy (mandate, capability registry, constraints, escalation) plus a post-hoc provenance audit is consistent with that cost structure, without claiming to derive it. Third, it instantiates every quantity it prices in a live institution, with the limits of that instantiation stated: one human-supervised hub dispatching to free-tier workers produced this manuscript, and the dispatch script, instrument table, frozen turn archive, incident log, and the agent panel's reviewer transcripts are public. Of the four layers, the mandate and needle-constraint layers ran; the registry and the escalation ladder are specified but did not run. The instrument logged 16 rows (13 briefs, 3 connectivity probes), all of which completed; the needle outcome for those rows is hub-attested, not logged, the dispatcher's own needle test could be satisfied by the CLI's echo of the brief, and the paper's own incident record shows at most 11 of 13 briefs were accepted as delivered on the first attempt. The hub-audited admission fraction ranged over [0.09, 0.19] across four waves. Two incidents, a worker that wrote outside its mandate under a variant of its needle and a fabricated worked example that passed its needle and two rounds of hub-briefed, non-blind agent review, were caught by provenance audit rather than by the escalation ladder. Three further incidents, all hub-side number errors, were caught by the agent reviewers and missed by the audit. That is the paper's practical lesson: needles are not acceptance, consistency-checking review does not stop fiction that fits the frame, and no single check was sufficient, since the hub's audit, its agent reviewers, and a second panel each supplied or missed an error that another check, or the same panel recomputing from the record, caught. What held was a reproducible source record and independent recomputation from it. Companion repository (instrument, data, incident log, reviewer transcripts, pre-registered wave): https://github.com/nKOxxx/agentic-institutions, tag v1.9.2.