LLM Agent Classrooms in Education

• Classroom of LLM agents is a multi-agent educational configuration where specialized language model agents collaboratively execute teaching, tutoring, and management tasks.
• Systems leverage role specialization, shared state, and structured communication to enable real-time instruction, personalized feedback, and simulation of pedagogical strategies.
• Empirical evaluations indicate that multi-agent classrooms enhance student engagement and adaptive learning, outperforming traditional one-on-one tutoring models.

Searching arXiv for papers on multi-agent LLM classrooms and educational agents. arXiv search query: "multi-agent LLM education classroom tutoring agents" A classroom of LLM agents is a multi-agent educational configuration in which multiple specialized LLM agents jointly perform teaching, tutoring, classroom management, peer interaction, assessment, content generation, analytics, or simulation functions within a shared instructional setting. Across recent systems, the classroom is instantiated either as a live instructional environment with teacher, assistant, peer, and coordinator agents; as a tutoring architecture with specialist feedback agents and persistent learner state; as a teacher-facing material-generation pipeline; or as a simulation platform for studying pedagogy, social dynamics, and adaptive scaffolding (Zhang et al., 2024, Yu et al., 2024, Elhaimeur et al., 27 Apr 2026, Ye et al., 28 May 2026). The common premise is that single-agent tutoring is often insufficient for heterogeneous educational objectives, whereas role specialization, orchestration, shared state, and structured communication can support scalability, adaptivity, interpretability, and richer social learning conditions (Chu et al., 14 Mar 2025, Cohn et al., 1 Feb 2026, Kumar et al., 3 Apr 2026).

1. Concept and scope

The term encompasses several closely related paradigms. In one line of work, a classroom is a live instructional setting populated by role-specialized agents such as a Teacher Agent, Teaching-Assistant Agent, AI classmates, and a hidden manager or session controller that selects which agent acts next (Yu et al., 2024, Zhang et al., 2024). In another, it denotes a multi-agent tutoring system whose agents are organized around specialist competencies such as video grounding, code guidance, conceptual scaffolding, autograding, learner profiling, or engagement monitoring (Elhaimeur et al., 27 Apr 2026, David et al., 21 Dec 2025). A third usage refers to teacher-facing systems in which multiple agents collaboratively generate instructional materials, individualized worksheets, or lesson plans (Gonnermann-Müller et al., 15 Aug 2025, Yao et al., 27 Aug 2025, Zhang et al., 7 Apr 2025). A fourth usage treats the classroom as a simulator populated by growable student agents and adaptive teacher agents for studying instructional interventions and social dynamics before deployment with real learners (Ye et al., 28 May 2026).

Survey work formalizes this broader landscape by distinguishing “Pedagogical Agents,” which automate pedagogical tasks for teachers and students, from “Domain-Specific Educational Agents,” which target fields such as science education, language learning, and professional development (Chu et al., 14 Mar 2025). Within pedagogical agents, the surveyed categories include Teaching-Assistance Agents, Student-Support Agents, Knowledge-Tracing Agents, and Error-Correction/Detection Agents (Chu et al., 14 Mar 2025). This suggests that a classroom of LLM agents is less a single architecture than a design family defined by multi-agent orchestration over educational tasks.

A recurring contrast is with one-on-one tutoring. Work on social learning explicitly argues that most AI-based educational tools adopt a one-on-one tutoring paradigm, whereas multi-party interaction can support peer modeling, co-construction, and exposure to diverse perspectives that dyadic tutoring alone cannot provide (Kumar et al., 3 Apr 2026). Controlled studies further show that participants interacting with both a tutor and peers achieved the highest unassisted test accuracy in a convergent problem-solving setting, while dual-agent assistance in writing avoided the idea-level homogeneity associated with single-model assistance (Kumar et al., 3 Apr 2026).

2. Core architectural patterns

Despite variation across applications, several architectural motifs recur. A basic pattern is role specialization plus orchestration. SimClass defines a set of agents $\mathcal{A}=\{\mathbf{a}_0,\dots,\mathbf{a}_N\}$, each instantiated by role customization, and a Session Controller composed of a Class State Receptor, a Manager Agent, and a library of pluggable functions (Zhang et al., 2024). The manager implements a decision rule

$$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$

selecting both the acting agent and the next tutoring or interaction function on the basis of the current class state (Zhang et al., 2024). MAIC uses a related Session Controller with a Class State Receptor and hidden Manager Agent that maps state $\mathcal{S}_t$ to an acting agent and teaching action $\mathcal{T}$ (Yu et al., 2024).

Another pattern is hub-and-spoke or spoke-and-wheel specialization. ITAS organizes its teaching layer as three parallel specialist agents—Video, Code, and Guidance—followed by a Synthesizer, plus a separate Autograder (Elhaimeur et al., 27 Apr 2026). Its response-time structure is explicitly decomposed as

$$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$

which reflects a parallel fan-out followed by serial synthesis (Elhaimeur et al., 27 Apr 2026). IntelliCode adopts a different but related design: a StateGraph Orchestrator coordinates six specialized agents for skill assessment, learner profiling, graduated hinting, curriculum selection, spaced repetition, and engagement monitoring, all operating over a centralized, versioned learner state under a single-writer policy (David et al., 21 Dec 2025).

Shared state is central in the more technically mature systems. EDU-oriented tutoring architectures differ in whether state is ephemeral, session-level, or persistent. IntelliCode uses a centralized Learner State containing mastery, uncertainty, misconceptions, reviews, engagement, preferences, and long-term memory, with every update schema-validated, atomically merged, versioned, and logged (David et al., 21 Dec 2025). AgentSchool instead models the entire simulation state as

$$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$

with local observations $o_{i,t}=\Omega_i(X_t,C_t)$ and global transitions $$X_{t+1}\sim\mathcal{P}_\theta(X_t,\{a_{i,t}\}_{i=1}^N)$$ (Ye et al., 28 May 2026).

Communication protocols also recur. The survey literature describes a prototypical orchestration in which a Teacher Agent receives student utterance $u_t$, broadcasts it to Student-Support Agents, and then uses a Coordinator or Critic Agent to aggregate tentative responses (Chu et al., 14 Mar 2025). Concrete implementations use request–response REST services (Gonnermann-Müller et al., 15 Aug 2025), JSON-typed inter-agent outputs enforced by Pydantic schemas (Elhaimeur et al., 27 Apr 2026), shared memory with asynchronous evidence-writing and synchronous decision reads (Cohn et al., 1 Feb 2026), or publish–subscribe buses and blackboard patterns in proposed classroom-scale designs (Fälton et al., 16 Jan 2026).

3. Agent roles and interaction regimes

Agent roles vary by educational objective, but the division of labor is usually explicit. In SimClass, the Teacher Agent carries TI, ID, EC, and CM responsibilities; the Assistant Agent supplements the teacher and maintains flow; and peer agents such as Class Clown, Deep Thinker, Note Taker, and Inquisitive Mind contribute humor, challenge, summarization, and questioning (Zhang et al., 2024). MAIC similarly includes a Teacher Agent, Teaching-Assistant Agent, AI Classmates, and future Recommendation/Analytics Agents (Yu et al., 2024). In the MAIC platform used for cluster analysis, six specialist agents are spawned per classroom session plus a director agent: AI Teacher, AI Teaching Assistant, Sparker, Questioner, Thinker, Note-Taker, and Director (Hao et al., 3 Mar 2025).

Other systems define roles around workflow stages rather than classroom personas. FACET comprises learner agents that simulate diverse profiles incorporating topic proficiency and intrinsic motivation, a teacher agent that adapts instructional content according to didactical principles, and an evaluator agent that provides automated quality assurance (Gonnermann-Müller et al., 15 Aug 2025). EduPlanner uses an evaluator agent, optimizer agent, and question analyst in an adversarial collaboration loop for mathematics lesson-plan generation (Zhang et al., 7 Apr 2025). Instructional Agents extends this model to teaching-faculty workflows via Teaching Faculty, Instructional Designer, Teaching Assistant, Course Coordinator, and Program Chair agents, aligned with the Analyze, Design, and Develop phases of ADDIE (Yao et al., 27 Aug 2025). ARISE, for hate-incident management training, uses a Manager Agent, three student agents with disciplinary personas, optional advisory agents, and a retrieval module (Gajewska et al., 30 Jun 2025).

A distinct interaction regime is adaptive scaffolding. EDF structures classroom support around three modules: Evidentiary Inference, Pedagogical Decision-Making, and Adaptive Feedback (Cohn et al., 1 Feb 2026). Its Evidence module is decomposed into StrategyAgent, AssessmentAgent, and KnowledgeAgent; its Decision module selects policies from $$\{\text{PROBE\_UNDERSTANDING},\text{SUGGEST\_ACTION},\text{PUSH\_LIMIT}\}$$; and its Feedback module instantiates the selected policy as a concrete talk move while implementing scaffold fading (Cohn et al., 1 Feb 2026). The rule-based policy can be written as

$$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$0

with thresholds tuned via pilot data and mapped to quintiles or deciles (Cohn et al., 1 Feb 2026).

Social-learning regimes depart still further from standard tutoring. In the convergent problem-solving study on SAT-level math, participants encountered Bob as a Socratic tutor, Alice as an arithmetic-error peer, and Charlie as a conceptual-error peer, with the tutor-plus-peers condition outperforming control and peers-only conditions on unassisted test accuracy (Kumar et al., 3 Apr 2026). In the divergent writing study, paired use of Claude and ChatGPT improved essay quality without increasing idea-level homogeneity relative to control, unlike the single-model condition (Kumar et al., 3 Apr 2026). The pedagogical mechanism is not formalized as a single theorem, but the paper’s design principles explicitly recommend combining authoritative tutor agents with fallible peer agents and engineering complementary error profiles to trigger schema conflict (Kumar et al., 3 Apr 2026).

4. Learner modeling, personalization, and adaptation

Personalization mechanisms differ substantially in granularity. FACET uses a persistent learner agent parameterized by $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$1 for topic proficiency and $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$2 for intrinsic motivation, encoded in prompt templates (Gonnermann-Müller et al., 15 Aug 2025). The summary formalizes the learner as sampling from $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$3 and sketches a logistic success model

$$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$4

along with a self-efficacy score $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$5 (Gonnermann-Müller et al., 15 Aug 2025). The Teacher Agent then selects task difficulty by maximizing a utility

$$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$6

over $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$7 (Gonnermann-Müller et al., 15 Aug 2025).

IntelliCode provides a more explicit centralized learner model. On each submission, mastery is updated per skill by a BKT-inspired rule:

$$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$8

followed by hint and time penalties and momentum smoothing (David et al., 21 Dec 2025). Curriculum selection then applies a 40/50/10 rule over due reviews, growth-zone skills with $$\mathcal{L}:\mathcal{S}_t\longrightarrow(\mathbf{a}_t,f_t),$$9, and challenge skills with $\mathcal{S}_t$0, subject to prerequisite filtering on a DAG (David et al., 21 Dec 2025). Spaced repetition uses SM-2 updates for ease factor and review intervals (David et al., 21 Dec 2025).

AgentSchool pushes learner modeling beyond prompt-conditioned personas by treating students as stateful learners with episodic memory $\mathcal{S}_t$1, a weighted subject knowledge graph $\mathcal{S}_t$2, thinking-workflow pools $\mathcal{S}_t$3, explicit misconceptions $\mathcal{S}_t$4, stable learner attributes $\mathcal{S}_t$5, and learning-style parameters $\mathcal{S}_t$6 (Ye et al., 28 May 2026). Its mastery updates are bounded:

$\mathcal{S}_t$7

separating latent cognitive state from observable performance (Ye et al., 28 May 2026). Teacher agents compute a ZPD-aligned scaffolding score

$\mathcal{S}_t$8

and select actions that balance group-level objectives and individual readiness (Ye et al., 28 May 2026).

At the classroom level, learner heterogeneity has also been modeled empirically rather than procedurally. In the MAIC cluster-analysis study, 110 university students were grouped into active questioners, responsive navigators, and silent listeners using agglomerative hierarchical clustering with Ward’s linkage on 23 features spanning interaction metrics, coded behavior/cognition/emotion proportions, Big Five traits, SRL, self-efficacy, attitudes toward AI, and pre-test score (Hao et al., 3 Mar 2025). The active questioners showed high messaging counts and elevated SC2/SC3, the responsive navigators showed more monitoring/regulation and off-topic content, and silent listeners produced zero messages while matching the other clusters on final-exam performance when controlling pre-test (Hao et al., 3 Mar 2025). This suggests that direct chat activity should not be equated with learning outcome.

5. Evidence, evaluation, and empirical findings

The literature evaluates classroom-scale LLM systems along several axes: interaction quality, learning outcomes, interpretability, teaching-material quality, operational reliability, and simulated-state validity.

SimClass evaluates interactional structure through the Flanders Interaction Analysis System and user experience through Community of Inquiry. In two real-world courses, removing all interactions dropped each CoI dimension by more than 30% with $\mathcal{S}_t$9, and removing classmates lowered Cognitive Presence and Social Presence significantly with $\mathcal{T}$0 (Zhang et al., 2024). Average user utterance length increased from 13.9 words in the “w/o classmates” condition to 18.9 words in full SimClass, while Student Talk doubled relative to the no-interaction ablation and Social Presence increased by about 0.4 points on a 0–2 scale when classmates were active (Zhang et al., 2024).

MAIC reports pilot data from Tsinghua University with more than 500 students and more than 100,000 learning records across two courses over approximately three months (Yu et al., 2024). Regression controlling for Module 1 score found that $\mathcal{T}$1 and $\mathcal{T}$2 positively predicted module score with coefficients significant at $\mathcal{T}$3 (Yu et al., 2024). Mode-usage analysis showed that 74% of students chose “continuous mode,” 26% preferred “interactive mode,” 61% of in-class student utterances were self-initiated questions, and 11% of student actions directed the agent, such as requesting the previous slide (Yu et al., 2024). Survey results included “Course Objective” mean 4.12 (SD 0.66), “Explore New Concepts” mean 4.03 (SD 0.73), and “Understand Student” 3.51 (SD 0.94) (Yu et al., 2024).

The MAIC cluster-analysis study contributes a different empirical layer. It analyzed 13,855 dialogue turns, about 1.6 million Chinese characters, across six modules, with LLM-assisted coding reaching 97% reliability after human review (Hao et al., 3 Mar 2025). The cluster findings sharpen the notion that multi-agent classrooms elicit heterogeneous participation styles rather than a single engagement pattern (Hao et al., 3 Mar 2025).

EDF/Copa is distinctive for measuring interpretability across the full Evidence→Decision→Feedback chain. It defines KeywordRecall for Data→Evidence grounding, SBERT similarity for Evidence→Decision alignment, and SBERT similarity for Decision→Feedback faithfulness, with significance established via permutation tests using 100–1,000 shuffles at $\mathcal{T}$4 (Cohn et al., 1 Feb 2026). In classroom deployment with $\mathcal{T}$5 sophomore dyads, 7,017 environment actions and 238 student-agent chat turns were collected; adaptivity was reflected in correlations between policy frequency and mastery quintile, including $\mathcal{T}$6, $\mathcal{T}$7, $\mathcal{T}$8, $\mathcal{T}$9, and $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$0, $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$1 (Cohn et al., 1 Feb 2026). KeywordRecall improved from 0.21 in the baseline to 0.43, SBERT(Evidence→Decision) from 0.39 to 0.64, and SBERT(Decision→Feedback) from 0.24 to 0.48, all with $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$2 (Cohn et al., 1 Feb 2026).

Controlled evidence for social-learning benefits comes from “Beyond the AI Tutor.” In convergent problem solving ($$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$3), mean unassisted test accuracy was 42% in control, 48% with peers only, 59% with tutor only, and 65% with tutor plus peers (Kumar et al., 3 Apr 2026). Pairwise $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$4 tests showed Tutor vs. Control $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$5, $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$6, and Tutor+Peers vs. Control $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$7, $$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$8 (Kumar et al., 3 Apr 2026). In divergent composition ($$L_{\text{total}}=\max\bigl(L_{\text{video}},L_{\text{guidance}},L_{\text{code}}\bigr)+L_{\text{synth}},$$9), essay-quality means were 2.32 for control, 2.65 for single-LLM assistance, and 2.55 for the dual-agent condition; idea homogeneity was 0.735 in control, 0.748 in single, and 0.737 in dual, with the dual-agent condition not differing from control on homogeneity (Kumar et al., 3 Apr 2026).

Teacher-facing generation systems use distinct evaluation protocols. FACET reports ten internal evaluations with mean suitability approximately 5.0/6 (SD approximately 0.06), mean clarity approximately 4.8/6 (SD approximately 0.12), and stability across runs with SD below 0.3 on all rubrics; external feedback from five K–12 teachers yielded didactical structure $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$0, suitability $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$1, clarity $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$2, and creativity $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$3 (Gonnermann-Müller et al., 15 Aug 2025). Instructional Agents adapts the Quality Matters rubric into six 1–5 component scores and reports human-rated overall quality $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$4 in Autonomous mode, $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$5 in Catalog-Guided, $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$6 in Feedback-Guided, and $$X_t=\bigl\{\{x^s_{i,t}\},\{x^t_{j,t}\},C_t,G_t,H_t\bigr\},$$7 in Full Co-Pilot mode, averaged over five CS courses (Yao et al., 27 Aug 2025). EduPlanner evaluates lesson-plan quality with CIDDP—clarity, Integrity, Depth, Practicality, and Pertinence—and reports experimental gains on GSM8K and Algebra, though the provided summary does not reproduce the exact aggregate dataset scores (Zhang et al., 7 Apr 2025).

Operational evidence is emphasized by ITAS. In a graduate quantum computing course, the teaching layer handled 334 chat turns, the operational layer captured 10,628 events across five modules, and the feedback layer surfaced two findings the instructor acted on mid-semester (Elhaimeur et al., 27 Apr 2026). The system recorded 7,666 video playback events, 387 code executions, a 77% code execution success rate, 32 checkpoint evaluations, and zero system-wide failures over the semester; end-to-end response latency was approximately 3.5–4 s under 1–50 concurrent users (Elhaimeur et al., 27 Apr 2026).

6. Design tensions, safety, and future directions

Several limitations and tensions recur across the literature. One concerns pedagogical principle versus learner expectation. In the EDF classroom study, students valued questions that “made me think” but sometimes found the system “broken” when it refused to give direct answers; teachers observed tension between principled scaffolding and expectations shaped by ChatGPT, and suggested upfront orientation plus in-dialogue transparency when students become stuck in loops (Cohn et al., 1 Feb 2026). A related implication is that multi-agent design must communicate not only content but also instructional role.

Another tension is between adaptivity and coherence. Proposed classroom-scale designs based on immersive visualization identify coherence as a problem when overlapping agent replies occur, recommending a centralized Dialogue Orchestrator and token passing (Fälton et al., 16 Jan 2026). The same design notes latency on the order of approximately 500 ms for speech-to-text, approximately 200 ms for the LLM, and approximately 100 ms for rendering, with local on-premise deployment suggested as a mitigation (Fälton et al., 16 Jan 2026). ITAS addresses related issues through typed JSON outputs, explicit task boundaries, and narrow-scope specialist agents, motivated by failures in an earlier prototype (Elhaimeur et al., 27 Apr 2026).

Safety, privacy, and fairness appear as cross-cutting constraints. The survey literature lists privacy and security risks from long-term memory, bias and fairness risks from stereotype reinforcement, hallucination and over-reliance risks from plausible but incorrect content, and integration risks stemming from inconsistent workflows and digital divides (Chu et al., 14 Mar 2025). Proposed mitigations include privacy-preserving memory, selective forgetting, on-device inference, fairness-checking classifiers, adversarial debiasing, retrieval-augmented grounding, teacher-in-the-loop review, and standard LTI interfaces (Chu et al., 14 Mar 2025). In classroom-specific designs for children, misinformation risk is addressed by proposing a fact-checker agent with a curated knowledge base, group-dynamics imbalance by a Reflection Agent that monitors turn counts and prompts quieter pupils, and child safety/privacy by secure audio streams with no user data persisted beyond the session (Fälton et al., 16 Jan 2026). ITAS implements pseudonymization at ingestion and keeps analytics within the institutional GCP environment as part of FERPA compliance by design (Elhaimeur et al., 27 Apr 2026).

The literature also repeatedly treats current results as preliminary. ITAS explicitly states that its pilot should be interpreted as system-behavior evidence rather than learning-outcome evidence and does not claim generalization from five students in one course (Elhaimeur et al., 27 Apr 2026). AgentSchool is motivated by the claim that real-world educational AI validation is slow, ethically constrained, and institutionally locked, so simulation is positioned as a research instrument rather than a substitute for all classroom trials (Ye et al., 28 May 2026). Several papers identify open questions around group-level dynamics, richer learner profiles, online updating of student models, and longitudinal effects on STEM interest or learning gains (Gonnermann-Müller et al., 15 Aug 2025, Cohn et al., 1 Feb 2026, Fälton et al., 16 Jan 2026).

A plausible synthesis is that the field is converging on three nonexclusive trajectories. The first is live classroom orchestration, exemplified by SimClass, MAIC, ITAS, and Copa, where agents interact directly with learners in real time (Zhang et al., 2024, Yu et al., 2024, Elhaimeur et al., 27 Apr 2026, Cohn et al., 1 Feb 2026). The second is instructional production for educators, exemplified by FACET, Instructional Agents, and EduPlanner, where the “classroom” is a coordinated design staff (Gonnermann-Müller et al., 15 Aug 2025, Yao et al., 27 Aug 2025, Zhang et al., 7 Apr 2025). The third is simulation and theory-testing, exemplified by AgentSchool and multi-agent peer-learning studies, where classrooms are used to model or experimentally probe adaptive pedagogy and social learning mechanisms (Ye et al., 28 May 2026, Kumar et al., 3 Apr 2026). Across all three, the defining technical challenge is not merely making agents speak, but making specialized agents coordinate over shared educational state in a way that remains interpretable, auditable, pedagogically aligned, and operationally robust.