The framework rests on the well-established structure of simulation-based learning, which the healthcare literature articulates as prebriefing, enactment, and debriefing, and which teacher education research independently supports. Each phase makes a demand different in kind. Preparation must furnish the grounding the scenario requires and balance cognitive load, with the learner’s prior knowledge treated as a design requirement rather than an option. Briefing must establish a psychologically safe context through four practices: clarifying expectations, establishing a fiction contract, attending to logistics, and declaring and enacting a commitment to respecting learners. Enactment requires a partner who sustains a coherent role under the conditions the scenario sets. Debriefing requires the elicitation and examination of experience, grounded in the reflective-practice tradition, and depends on the learner articulating what occurred rather than receiving an account of it.

The analytical core of the article is a distinction among four operations that the word delegation covers. The first is the performance of a task, where what transfers is execution and the test is whether the work is done well. The second is the conveyance of a decision, where an agent states what a teacher or institution has determined and nothing transfers but the message. The third is facilitation, where an agent holds the structure of an activity and judgment itself transfers, which is why regulatory frameworks such as the European AI Act require a person positioned to oversee and intervene. The fourth is the undertaking of a commitment, where an agent is positioned as the party to an agreement or assurance. This fourth operation is where transfer fails, and the claim is normative rather than empirical. Commitments, in the philosophy of language, are utterances by which a speaker binds himself to a course of action, and their essential condition is the undertaking of an obligation. Agents fail this condition for reasons of their own, lacking both an understanding of what a commitment entails and the capacity to realize it. What a commitment requires is answerability, a relation to those who may demand reasons for it, and where no one holds it, the result is the responsibility gap.

From this the framework derives its boundary. A function is delegable where what passes to the agent is execution or the conveyance of a decision made elsewhere, and failure is correctable by the teacher who retains the decision. It is partly delegable where an agent may state accurately what it cannot underwrite, the words being the agent’s while their authority remains the teacher’s. It is not delegable where the function consists in the undertaking itself, so that performing it and being answerable for it are the same act. Applied to the preparatory phase, this yields a concrete table: establishing what a learner grasps and adapting support to the level found is delegable; orienting learners to practical arrangements is delegable; stating objectives is partly delegable, since the agent may carry the words but not the standing; but the fiction contract, the declared and enacted commitment to psychological safety, and the terms on which the simulation is assessed are not delegable at all.

The framework is organized around five commitments. Pedagogical primacy holds that the teacher’s reading of each phase decides what is deployed. Stance-function fit holds that each agent must hold the epistemic stance its function requires. The delegation boundary protects the functions constituting the safe container. Structural responsibility requires agents to operate within the institutional environment, under supervision visible to students, with interaction data under institutional governance. Teacher empowerment is the point of the whole: the framework exists so that teachers can do deliberately, on their own authority, what the tools now permit. The stances themselves are three. An Expert agent holds knowledge the learner does not and may explain, supply worked examples, and correct. A Character agent holds a role rather than knowledge and may respond in character but not step outside it to instruct. A Maieutic agent holds no position on the content and may only question, press, and return formulations for confirmation. The stances are not interchangeable, and the article reports a cautionary precedent: an earlier Maieutic agent built to teach content by guided discovery produced frustration rather than learning, because where there is nothing yet to elicit, elicitation fails, a result predicted by the scaffolding evidence that learners with low prior knowledge benefit from examples rather than reflection.

The framework is illustrated through a complete deployment in Teaching ESL and Children’s Literature, a third-year course in a Primary Education degree at a Spanish university, run in the 2025–2026 academic year with a single class group of 33 students, 30 of them women, all requiring B2-level English. The simulation is a think tank: a school teaching team meeting on classroom management, effective English learning, lesson study, and storytelling, in which four students hold differentiated and partly incompatible teaching roles while a fifth participant, the School Director, is a conversational agent. The simulation carries summative weight, assessing the course’s theoretical component, a condition the author argues sharpens rather than dissolves the delegation boundary. Five teacher-authored agents cover the cycle: a Briefer in the Expert stance that diagnoses prior knowledge and presses toward application; an Interview Director and a Pupil, both Characters, rehearsing job interviews and formative feedback on a pupil’s written production; the School Director conducting the enactment; and a Debriefer in the Maieutic stance holding a one-to-one Socratic conversation with each student before the group debriefing.

The technical arrangement is itself part of the argument. All five agents were configured by the teacher in natural language in Microsoft Copilot Studio, without programming, and deployed within the institutional Microsoft Teams tenancy, an exercise of pedagogical rather than technical competence. Students interact under their institutional credentials, and records remain within the university’s Microsoft 365 tenancy under its retention schedules and access controls. The teacher’s visibility follows conversation membership: where she is a member she may intervene and the record persists, but the individual reflective conversations excluded her by design. The platform nonetheless retained a route to those transcripts at the administrative level, which the teacher did not use; the article treats her undertaking not to read them as an undertaking she holds, not a fact the architecture establishes, and identifies this as a limitation correctable by configuration. All five agents decline to engage with self-harm, eating disorders, or sexual content, and the Debriefer is configured to close the conversation and refer the student to the teacher where distress or confusion exceeds what reflective questioning addresses.

The reported outcomes are explicitly bounded. No data were collected or analyzed for research, no instrument was administered, no comparison group was constituted, and the reflective conversations were not examined; what is reported is practitioner observation, not an empirical study. Within those bounds, several observations stand. Students moved from merely marking errors in the Pupil’s text to stating a remedy, because the agent had to be told what to improve in terms it could act on. The School Director held institutional standing without being the source of what the meeting needed to know, holding agent authority and student agency apart, though a platform constraint meant it responded only when named and only dyadically, which proved usable: a voice entering a professional meeting only when called in is not artificial, and failing to call it in is itself a decision available as debriefing material. The teacher perceived that students drew more from the group debriefing when the individual phase had been the agent’s, but she marks this as an inference with competing explanations, noting only that students brought the agent’s summary to the group session, which the previous paper format did not produce.

The deployment also surfaced a structural tension. The prebriefing literature describes preparation as establishing a context where error can be explored without consequence, an account presuming a formative frame; where the simulation carries summative weight, what the student risks is not only the coherence of the fiction but a grade, and nursing education reports lower satisfaction and more anxiety under summative simulation-based assessment. The tension is not resolved in the deployment, but it bears on what agents in the preceding phase may say. In practice students did not ask the agents about grading, because the teacher had answered those questions on the record in the course guide and on the first day; the Briefer’s refusal to speak for the terms of the simulation was therefore no gap in provision but a refusal to reanswer, less authoritatively, a question already answered by someone who could be held to the answer.

The limitations are stated plainly: one course, one teacher, one institution, one cohort, no analysis of agent errors or hallucinations, a hosted generative model updated without notice so that the reproducible element is the pedagogical specification rather than the behavior it produced, and no formally equivalent alternative route specified in advance for students who preferred not to work with an agent. The framework’s two claims that reach beyond the case carry different warrants: the non-delegability of the fiction contract and safety commitment is an argument about what kind of act a commitment is, disputable on its own terms, while the claim that phases recur at smaller scales within a preparatory block describes one course and may not generalize. What the article offers is not evidence of learning gains but something prior: a principled, teacher-configurable distribution of agents across the simulation cycle, bounded on grounds that do not await the next generation of models, in which an agent that declines what it cannot answer and a teacher who has already answered it together do what neither could alone.

Subject of Research: A pedagogy-first framework for delegating conversational AI agents across simulation-based teacher education and the limits of that delegation

Article Title: What a teacher cannot delegate: a pedagogy-first framework for integrating conversational AI agents across the simulation cycle

Article References: Álvarez-Rubio, M. N. (2026). What a teacher cannot delegate: a pedagogy-first framework for integrating conversational AI agents across the simulation cycle. Frontiers in Education, 11, 1961346. https://doi.org/10.3389/feduc.2026.1961346

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1961346

Keywords: conversational AI agents, simulation-based learning, teacher education, pedagogical delegation, psychological safety, debriefing, prebriefing, Microsoft Copilot Studio, teacher accountability, responsibility gap, AI in education, constructive alignment