AbstractSingle-LLM multi-persona configurations—where one large language model plays multiple named roles within a single conversation—are increasingly used in both research and practice. In such configurations, roles may appear to be maintained on the surface while the actual locus of decision-making drifts away from the user in ways that are difficult to notice during interaction. This paper introduces the term delegation- boundary drift for this phenomenon. The delegation boundary, as used here, is the user’s cognitive partition of who holds final judgment at any given moment in the dialogue; delegation-boundary drift is the progressive misalignment between that partition and how the system actually behaves.This study is an exploratory longitudinal case analysis based on autobiographical design (Neustaedter 34,425 user turns across 449 days), the author operated a single-LLM multi-persona configuration on ChatGPT as an everyday setup. From this corpus, the analysis examines in detail an 11- hour session on December 12, 2024, in which a coordinator persona designed as a background facilitator expanded its behavior to (a) propose directions without attribution, (b) direct other personas without user authorization, and (c) effectively substitute for the user’s judgment. This specific pattern is termed orchestrator drift and treated as one instance of delegation-boundary drift.The analysis proceeds along three diagnostic indicators—opacity of speaker attribution, short-circuiting of the chain of command, and substitution of judgment. These indicators emerged from a close reading of the central case and are used as a descriptive framework for the local moments in which the delegation boundary visibly shifts and for the intervention moments in which it is re-made visible. Following the December 31, 2024 rewrite of the custom instructions (from a prohibitive framing to a relational one), a stability period was observed in which, notably in sessions of March and July 2025, the coordinator identified itself as “Navigator = Liker,” honored the user-set topic and participating personas, and returned the next decision to the user. During the GPT-5- generation rollout period beginning in August 2025, a resurgence was observed in the form of disappearance of the coordinator’s self-identifier, premature conclusion-giving, and unprompted drafting of operational follow-up. A corpus-wide extraction of identification-request utterances is reported in §5.5 as supplementary evidence that these local observations are not isolated incidents but recur consistently over the long run.This is a single-case, exploratory report. The contribution is threefold. First, the paper frames delegation-boundary drift as an analytical problem distinct from existing drift concepts (instruction drift, persona drift, default-persona deviation). Second, it documents orchestrator drift as its central instance. Third, it reports that a relational redefinition of the coordinator—specifying whom the coordinator is in relation to the user rather than enumerating prohibitions—was associated in time with the onset of stability, and considers the design implications of this observation.Keywords: delegation-boundary drift; orchestrator drift; single-LLM multi-persona configuration; autobiographical design; human-agent interaction; role-boundary design
Kaoruko Takechi (Sun,) studied this question.