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

Synthesis: Álvarez-Rubio argues that conversational AI agents can carry much of a Simulation cycle and still not all of it, because the line is drawn by what kind of act a function is rather than by what current technology can do. Five commitments follow: the teacher's reading of each phase decides deployment, an agent's stance must fit the phase's demand, three functions are constitutively non-delegable, supervision must be visible to students, and the point is teacher authority rather than replacement. An agent's most valuable move at the boundary is refusal.

Conceptual paper: this is a practitioner design inquiry, not an empirical study. The author states that "no data were collected or analyzed for research" and that the observations are her own, so the deployment illustrates the framework without testing it.

What the framework argues

1. Pedagogical primacy. The teacher's reading of each phase's demand decides what is deployed, where, when, and whether. No catalogue of agents substitutes for it.
2. Stance-function fit. Each agent holds the stance its function requires and draws on the knowledge of the course it serves. The stances divide by relation to knowledge, not to the learner.
3. Delegation boundary. The functions that constitute the safe container are not candidates for integration at all. This is the contribution, and it is normative rather than empirical.
4. Structural responsibility. Agents operate inside the institutional environment, under supervision visible to students, with interaction data under institutional governance.
5. Teacher empowerment. The framework exists so that teachers can do deliberately, on their own authority, what the tools now permit.

Where the boundary falls, and why it falls there

Delegation covers four operations. Execution passes work to the agent, and the test is whether it is done well. Conveyance passes only a message. Facilitation passes judgment over the structure of an activity. The fourth is the undertaking of a commitment, where the agent is positioned as the party to an agreement.

That fourth is where transfer fails. A commitment is an utterance by which a speaker binds himself to a course of action, and its essential condition is the undertaking of an obligation. It misfires where the speaker cannot undertake it, and agents fail twice: they lack an understanding of what a commitment entails and the capacity to realize it. Responsibility for machine speech traces to persons, institutions, or no one.

The decisive move separates performing a function from answering for it. Answerability is a relation to those who may demand reasons, not a property of an utterance, and where nobody holds it the responsibility gap opens. Delegable functions pass execution or conveyance and let the teacher correct a failure. Partly delegable ones let the agent state what it cannot underwrite. Non-delegable ones are where performing the function and answering for it are the same act.

Applied to preparation, that yields six verdicts. Diagnosing what a learner grasps and supplying what is missing is delegable, since collective instruction cannot adapt to individual prior knowledge. Orienting students to logistics is delegable conveyance. Stating objectives is only partly delegable. The fiction contract, the psychological safety commitment, and the terms of assessment are not delegable at all.

The test travels as three questions: what would pass to the agent; whether the teacher could correct a failure; and whether anyone could be asked to stand behind the agent's words. A function with no answer to the third is not a candidate.

The cycle, and the smaller cycles inside it

The phases are constituted by demands rather than content, which is why the prebriefing standard travels from healthcare simulation into teacher education without carrying clinical assumptions. Preparation establishes the grounding the scenario requires, enactment requires a counterpart that sustains a role, and debriefing depends on the learner articulating what occurred, the reflective practice tradition the phase inherits. The phases also recur at smaller scales within a nominally single phase, which is why preparation here holds two agents of different stances.

Stances: expert, character, maieutic

An Expert agent holds knowledge the learner does not: it may explain, supply worked examples, and press toward application, but not withhold what the learner needs in order to elicit it. A Character agent holds a role rather than knowledge, and may respond in character, resist, and react, but not step outside the role to instruct or evaluate. A Maieutic agent holds no position on the content: it may question and return a formulation for confirmation, but not interpret, advise, or answer.

A preceding course tested the distinction. A Maieutic agent built to teach by never answering did not work: students were frustrated, because discovering what one does not know is difficult and a course is short. The Scaffolding evidence predicts the same failure: low prior knowledge needs examples, while high prior knowledge benefits from reflection.

The simulation that assesses

Here the simulation carries summative weight, which the paper treats as the hardest test of its own framework. The prebriefing literature describes preparation as a context where error can be explored without consequence, presuming a formative frame. Where the simulation also grades, the student risks a grade, and nursing education reports lower satisfaction and more anxiety under summative than under formative simulation assessment.

The tension belongs to the combination, not to any implementation, and it bears on what preceding-phase agents may say: the Briefer declines the terms of the simulation, refusing to reanswer what the course guide had settled.

The deployment: five agents in one course

The case is a third-year course, Teaching ESL and Children's Literature, in a Primary Education degree at a Spanish university. Thirty-three students took part in the 2025–2026 second semester, from February to the end of May 2026: 30 women and three men, including one mature student and five exchange students.

The simulation is a think tank, a school teaching team meeting on classroom management, English learning, lesson study, and storytelling. Four students hold differentiated roles composed so that sustaining one position requires arguing from the course theory, and a fifth participant, the School Director, is an agent. The Briefer, Interview Director, Pupil, School Director, and Debriefer cover the cycle, each holding one of the three stances set out above.

All five were authored by the teacher in natural language without programming, inside the institution's existing Teams tenancy. That is a condition of the argument: if alignment depends on reading what a phase demands, nobody who does not know the course can settle it. Supervision is visible through conversation membership, so a student can see who is present rather than taking anyone's word for it.

What the architecture does not settle is what the platform retains. Transcripts were written to a data table, and the teacher's not reading them is an undertaking she holds rather than a fact the architecture establishes.

What the deployment showed, and what it could not

Three questions can be asked of any distribution: whether each agent behaved as specified, whether the tasks produced the work the objectives called for, and whether the sequence changed learning. The first two are addressed from what was visible in class and in the teacher's conversations. The third is not addressed: no instrument, comparison group, or coding was used.

The Interview Director put difficult concrete situations to students from the first turn, and some did not reach an adequate response; those cases were taken up in a dialogic lecture where no answer was presented as correct. The Pupil made a narrower failure visible: students marked errors, but seeing symptoms did not produce an adjusted diagnosis.

The School Director is the most interesting negative result. It convenes the meeting on the premise that the students are the experts and then largely listens, holding standing without being the source of what the meeting needs to know. In a multiparticipant conversation it responds only when named, so its participation stays dyadic: in the group without being of it.

On debriefing, the teacher perceived that students drew more from the group session when the individual phase had been the agent's rather than written questions. She marks that as an inference, with the task new and herself the designer of the change. What is observable is narrower: students brought the agent's summary to the group session and used it there.

The Debriefer declares the conversation a space for thinking rather than an evaluation, and its closing summary belongs to the student and is not transmitted to the teacher, who is not a member. That absence is constitutive of the reflective contract. An agent cannot judge whether discomfort is productive or overwhelming, so it refers the student to the teacher.

A chatbot that debriefs as well as a human does not thereby become answerable, and that is what the boundary turns on.

What this means for practice

• Teachers authoring agents. Apply the test per function, not per phase. Ask what passes to the agent, whether you could correct a failure, and who could be asked to stand behind what it said.
• Teacher educators. Teach refusal as a feature. An agent that declines to speak to grading or psychological safety is doing what the framework asks, if the teacher has answered on the record.
• Program leaders. Expect the boundary to matter more where a simulation carries assessment weight, and specify in advance an equivalent route for a student who declines to work with an agent.
• Administrators. Inspect what the platform retains before accepting a promise about it. Architecture narrows what must be promised; it does not close the boundary.
• Researchers. Test whether stance-to-phase alignment changes what students learn. The comparison between agent-led and paper-based reflection is the strongest observation and most in need of measurement.

Limitations

• One course, one teacher, one institution, and a cohort of 33 students; no data were collected or analyzed, no reflective output was examined, and there was no comparison group.
• No analysis of agent errors or hallucinations was undertaken, and role drift was checked informally rather than tested.
• The agents run on a hosted model the vendor updates without notice, so what reproduces is the pedagogical specification rather than the behavior it produced.
• The Briefer's diagnosis of prior knowledge proceeds through conversation and was not validated against an independent measure.
• Two ethical arrangements rest on undertakings: the teacher's access to reflective conversations was governed by her promise, and no alternative route was specified in advance.
• The deployment ran in English; whether summative assessment altered use is unknown.

Citation

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.