Instructional Agents

• Instructional agents are a family of AI systems with explicit pedagogical, procedural, and design roles that support teaching, learning, and instruction execution.
• They range from conversational chatbots and teachable agents to multi-agent frameworks automating course design, live tutoring, and material generation.
• Key challenges include adaptive instructional behavior, reliable evaluation, and addressing safety and security issues in dynamic educational environments.

Instructional agents are AI systems whose operation is organized around instructional structure. In current research usage, the term spans several related classes of systems: pedagogical agents that support learner-centered education through scaffolding, feedback, and social interaction; LLM-based agents for Instructional Systems Design (ISD) that reason over the ADDIE lifecycle; teacher-facing systems that generate or adapt educational materials; and instruction-following agents that read manuals or external documentation as actionable guidance (Hare et al., 25 Aug 2025, Jeon et al., 11 Feb 2026, Wu et al., 2022). What unifies these lines is not a single interface form, but a common commitment to structured instructional activity: teaching, being taught, designing instruction, or executing instructions.

1. Scope and conceptual foundations

The contemporary literature does not treat instructional agents as merely conversational answer generators. In educational technology, an interactive pedagogical agent may be a screen-based “learning partner” that guides users through learning tasks; in one formulation, the agent is an LLM-powered conversational chatbot for instructors that provides teaching suggestions rather than fully autonomous tutoring (Chen et al., 6 Mar 2025). In ISD research, an instructional agent is a specialized class of LLM-based agent for end-to-end educational program design, judged not just on fluency or correctness but on whether it can produce professionally valid instructional design under realistic contextual constraints (Jeon et al., 11 Feb 2026). In course-development work, the agent may be a multi-agent workflow that automates syllabus creation, lecture scripts, LaTeX-based slides, and assessments while preserving the coordination normally distributed across teaching faculty, instructional designers, TAs, and program leadership (Yao et al., 27 Aug 2025).

A further extension broadens the term beyond education-as-content-delivery to education-as-actionable procedure. Research on action condition inference argues that the kinds of knowledge needed by autonomous instruction-guided agents and assistive AI include preconditions and postconditions of actions, not only ordered step lists (Wu et al., 2022). At the opposite end of the spectrum, security-oriented work studies high-privilege LLM agents that read project documentation, reason about what to do, and act through a shell, filesystem, browser, or network; these systems are instructional agents in the literal sense that they treat external natural-language documentation as instructions (Kao et al., 12 Mar 2026).

This breadth has a methodological consequence. A recurrent theme across the literature is that instructional agency is better understood as a family resemblance among systems with explicit pedagogical, procedural, or design structure than as a synonym for “chatbot.” A plausible implication is that taxonomies of instructional agents are most informative when organized by instructional role—teacher, learner, designer, assistant, or executor—rather than by interface alone.

2. Learning-by-teaching, affect, and the teachable-agent tradition

One foundational lineage defines the teachable agent as a special type of pedagogical agent that instantiates the educational theory of Learning by Teaching (Borjigin, 2015). In this paradigm, the learner becomes the tutor and the agent becomes the tutee. The reported pedagogical rationale is that students organize knowledge, take responsibility for it, and reflect on and explain it more deeply when they teach. Early critiques of teachable agents in this line emphasized two deficits: lack of proactive interaction and lack of believability sufficient to evoke empathy and sustain engagement (Borjigin, 2015).

The Affective Teachable Agent (ATA) was proposed to address those deficits through a goal-oriented architecture that couples Teachability and Affectivability. Its formalization is given as

$ATA = (E, P, K, G, Rt, A),$

with $Rt = (Tr, Ar)$, where $Tr$ denotes Teachability Reasoning and $Ar$ denotes Affectivability Reasoning (Borjigin, 2015). Goal Net provides the hierarchical control mechanism through which the agent can pursue goals rather than respond passively. This move is significant because it makes proactive interaction an architectural property rather than a surface behavior.

The Persuasive Teachable Agent (PTA) extends the same learning-by-teaching logic by adding persuasion, social engagement, and intergenerational learning support (Lim, 2016). PTA combines the Elaboration Likelihood Model, Goal Net, and Fuzzy Cognitive Map reasoning. Its architecture is formalized as

$PTA = (E, P, K, Tr, Pr)$

and later as

$PTA = (E, ET, K, Rs, R, A),$

with $R = (P, Tr, Pr)$ (Lim, 2016). PTA generates peripheral persuasive cues drawn from ExpertHintsSet, AttractiveSourceSet, and AffectSet, with interface realizations such as facial expressions, speech bubbles, reminders, encouragement, hints, and emotional animations. This suggests a broader conception of instructional agency in which motivation and participation are not secondary user-experience issues but part of the instructional mechanism itself.

Conversational learning-by-teaching systems adapt the same principle to natural-language dialogue. Curiosity Notebook investigates conversational agents that support the Protégé effect by acting as learners taught by students in natural language rather than as conventional pedagogical agents that instruct students (Chhibber et al., 2019). The agent asks questions, requests examples or rephrasings, exposes partial understanding, and uses “cognitive, affective, and meta-cognitive interventions” to sustain engagement. In professional development, Novobo applies the same role reversal to teachers’ peer learning of instructional gestures: the AI is framed as a novice or mentee, teachers jointly critique its generated gestures, demonstrate alternatives through a skeletal mirror, and articulate the reasoning behind them (Jiang et al., 23 May 2025). A plausible implication is that the tutee-agent form is especially effective when the target knowledge is tacit, embodied, or socially delicate.

3. Procedural knowledge, goal inference, and adaptive instructional behavior

A distinct research thread treats instructional agents as systems that must understand what instructions require in order to act safely and correctly. Action condition inference formalizes this requirement as prediction over directed labels $$\{\text{NULL}, \text{precondition}, \text{postcondition}\}$$ between text segments extracted from instructional manuals (Wu et al., 2022). The task assumes the segments are given and asks the model to infer the dependency graph encoded by the manual. Empirically, contextualized modeling of the full instruction paragraph improved performance by more than 20 F1 points over a non-contextualized pairwise baseline, and heuristic weak supervision contributed more than 6 F1-point gain in the low-resource setting (Wu et al., 2022). This suggests that procedural instructional understanding is fundamentally discourse-level rather than reducible to local relation classification.

Goal-conditioned communication research reframes instructional interaction as goal inference rather than behavior copying. Goal-conditioned agents that function as both teachers and learners use action-based demonstrations and language-based instructions while incorporating pedagogy on the teacher side and pragmatism on the learner side (Caselles-Dupré et al., 2023). In the reported Fetch Block Stacking setting, Pedagogical Teacher + Pragmatic Learner outperformed Naive Teacher + Literal Learner in every modality. The best result, Demonstrations then Instructions, reached an AuC of 36.2, compared with 34.3 for the naive/literal pairing; demonstrations only improved from 28.8 to 35.7 (Caselles-Dupré et al., 2023). The operational lesson is explicit in the paper: teach goals, not just behaviors; make teachers pedagogical; make learners pragmatic; and use multiple modalities when possible.

More recent architectures split instructional support across specialized pedagogical roles. A neuro-symbolic framework assigns an RL-based Tutor Agent to authoritative, non-verbal scaffolding and an ontology-grounded LLM Peer Agent to proactive social support (Hare et al., 25 Aug 2025). The tutor policy is written as

$a_t^* = \pi^*(S_t),$

while the peer agent conditions generation on retrieved ontology knowledge $K_O$ and state $Rt = (Tr, Ar)$0 (Hare et al., 25 Aug 2025). The central educational ontology stores domain concepts, prerequisites, difficulty levels, associated media, pedagogical rules, and grounding facts. This decomposition matters because it treats technical support and social support as distinct but coordinated instructional functions.

Simulation work pushes the same logic toward adaptive teacher modeling. One framework evolves a teacher agent’s pedagogical policy with a genetic algorithm over explanation style, content focus, delivery pace, and engagement mode, using average student assessment score as fitness (Sanyal et al., 25 May 2025). Persona-RAG then conditions retrieval on a student-specific reasoning plan rather than on the question alone. In a different formulation, AgentSchool models learning as state transition rather than prompted behavior, with teacher agents that plan, scaffold, and reflect along the Zone of Proximal Development and student agents equipped with weighted subject knowledge graphs, thinking-workflow pools, and explicit misconceptions (Ye et al., 28 May 2026). The ZPD compatibility score is defined explicitly, and teacher effects are reported as backbone-dependent rather than universally dominant (Ye et al., 28 May 2026). A plausible implication is that adaptive instructional agency depends not only on pedagogical architecture but also on the representational fidelity of learner state.

4. Instructional design and educational material generation

Teacher-facing and designer-facing instructional agents have increasingly been formulated as multi-agent ISD systems rather than single-pass content generators. An LLM-based framework titled Instructional Agents automates end-to-end course material generation across the Analyze, Design, and Develop phases of ADDIE, with role-specialized agents for Teaching Faculty, Instructional Designer, Course Coordinator, Teaching Assistant, and Program Chair (Yao et al., 27 Aug 2025). It supports four operating modes—Autonomous, Catalog-Guided, Feedback-Guided, and Full Co-Pilot. Full Co-Pilot consistently yielded the best material quality, while Autonomous required the least human effort; with gpt-4o-mini, Autonomous averaged 1.46M tokens, 2.23 hours, 0 min human time, and $Rt = (Tr, Ar)$10.36 (Yao et al., 27 Aug 2025). The paper explicitly frames this as a quality–efficiency frontier.

Benchmarking work has turned ISD-oriented instructional agents into a measurable research problem. ISD-Agent-Bench comprises 25,795 scenarios generated through a Context Matrix with 51 contextual variables across 5 categories and 33 ISD sub-steps derived from ADDIE (Jeon et al., 11 Feb 2026). Evaluation combines ADDIE rubric assessment and trajectory evaluation with

$Rt = (Tr, Ar)$2

On 1,017 test scenarios, React-ADDIE achieved the highest overall score, 86.49, exceeding Baseline (84.07), Dick-Carey-Agent (84.20), RPISD-Agent (83.62), and ADDIE-Agent (82.96) (Jeon et al., 11 Feb 2026). The reported result that theoretical quality correlates with benchmark performance at $Rt = (Tr, Ar)$3 supports the claim that instructional agents benefit from classical instructional theory when paired with reflective agent architecture (Jeon et al., 11 Feb 2026).

Several systems move pedagogical expertise from the user prompt into the agent architecture itself. A KLI-guided study compares a Single-Agent baseline, a role-based sequential MAS, and a collaborative MAS-CMD using conquer-and-merge discussion to generate secondary Math and Science learning activities (Wang et al., 20 Aug 2025). Human QM totals were 10.05 for MAS-CMD, 9.55 for SAS, and 9.05 for MAS-Roles, with the clearest statistical difference on QM 5.2C Active Learning (Wang et al., 20 Aug 2025). Teachers nevertheless reported that MAS-CMD outputs were more creative, contextually relevant, and classroom-ready than either comparison system. FACET similarly defines a teacher-centered multi-agent pipeline—learner agent, teacher agent, evaluator agent—for individualized grade 8 mathematics worksheets that incorporate both topic proficiency and intrinsic motivation (Gonnermann-Müller et al., 15 Aug 2025). Its four prototype learner profiles are denoted

$Rt = (Tr, Ar)$4

and outputs are scored on didactical structure, clarity of the task, creativity, and suitability for learners’ needs on a 1–6 scale (Gonnermann-Müller et al., 15 Aug 2025).

Human-centered work for higher education instructors adds a complementary perspective. An interactive pedagogical agent for instructors was developed through interviews with five pedagogy experts and participatory design with ten pedagogy experts, then evaluated on 25 instructor questions (Chen et al., 6 Mar 2025). The prototype used retrieval-augmented generation over 115 active-learning examples from ABLConnect, retrieving the top three similar examples for each question (Chen et al., 6 Mar 2025). Experts preferred the prototype’s outputs over default ChatGPT, with about one point higher scores across all dimensions and about two points higher on Actionability. Trust-building features emphasized in this work—incremental and flexible information collection, peer validation, social transparency, conversational control, and privacy assurances—show that instructional design agents are not only content generators but also sociotechnical interfaces for professional judgment.

5. Live delivery, tutoring support, and embodied instruction

Instructional agents are increasingly deployed not only before instruction but during live teaching. One field deployment used an LLM-based Instructor Agent as the primary instructional interface in a graduate-level Cloud Computing course at IISc, integrated into Microsoft Teams Copilot with a base system prompt, pedagogical alignment prompts based on the Knowledge–Learning–Instruction framework, and a weekly topic knowledge base tied to Bloom’s taxonomy (Simmhan et al., 23 Oct 2025). Transcript analytics defined three interpretable engagement metrics—topic coverage, topic depth, and turn length—and early results across two modules showed coverage decreasing from about 52.5% to about 31.0%, while depth increased from 1.33 to 2.06 and turn length from 48.2 words to 54.4 words (Simmhan et al., 23 Oct 2025). The paper interprets this as a transition from broad conceptual exploration to deeper, focused inquiry.

Controlled evidence for live instructional delivery appears in a randomized controlled trial comparing a traditional human teacher, a self-paced MOOC with chatbot support, and an AI instructional agent capable of delivering lectures and answering questions in real time (Qin et al., 28 May 2025). After exclusions, the analytic sample comprised 125 students: 41 in the human teacher group, 43 in the MOOC group, and 41 in the AI instructional agent group (Qin et al., 28 May 2025). The AI group showed significantly higher perceived learner control, with F(2, 122) = 12.155, p < .001; more frequent interaction, with F(2, 122) = 21.755, p < .001; and higher post-test performance, with F(2, 122) = 7.673, p < .01 (Qin et al., 28 May 2025). Regression further showed perceived learner control predicting post-test performance with β = 0.055, p < .05. This suggests that instructional agency can be pedagogically consequential when pace control and real-time responsiveness are embedded in the same instructional identity.

Embodied teaching extends the delivery problem beyond text and speech. LectūraAgents proposes a hierarchical multi-agent framework in which a ProfessorAgent orchestrates ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, and TasaAgent, then delivers a personalized lecture through visible teaching actions such as highlighting, underlining, circling, boxing, and handwriting in a slide environment (Sesay et al., 15 Jun 2026). Its Teaching Action-Speech Alignment mechanism defines each action as

$Rt = (Tr, Ar)$5

with actions aligned to slide content, speech timing, learner profile, and pedagogical salience (Sesay et al., 15 Jun 2026). The system was evaluated on 280 lectures total across seven frontier models and four academic levels, with five expert educators assessing lecture content quality, personalization quality, assessment quality, and teaching action quality (Sesay et al., 15 Jun 2026). The paper reports particularly strong performance on spatial accuracy, handwriting, rough notation, and embodied teaching.

Real-time tutoring support can also take the form of material retrieval rather than lecture delivery. An on-demand instructional material providing agent based on a multimodal LLM listens to spoken tutor–student dialogue, generates exactly five search queries every 10 seconds using Google Gemini 1.5 Flash, and retrieves Web images for immediate instructional use (Kato et al., 26 Mar 2026). In evaluation, the agent reduced average image retrieval time from 65.7 ± 29.4 s without support to 21.3 ± 7.0 s with support, a reduction of 44.4 s, and supplied images of acceptable quality in 85.7% of trials (Kato et al., 26 Mar 2026). This widens the notion of instructional agency: the agent need not itself explain the concept if it can reduce the tutor’s interruption cost in finding the right explanatory resource.

6. Evaluation, reliability, and security constraints

Because instructional agents increasingly influence pedagogy, evaluation methodology has become a central research topic. ISD-Agent-Bench explicitly addresses LLM-as-judge bias through a multi-judge protocol using GPT-4o-mini and Gemini-2.5-flash-lite, with final scores aggregated by median (Jeon et al., 11 Feb 2026). Reported reliability is strong: the main text highlights mean reliability = 0.905, while the appendix reports ICC(2,k) = 0.891 overall and Krippendorff’s $Rt = (Tr, Ar)$6 overall (Jeon et al., 11 Feb 2026). This is important because instructional quality criteria such as coherence, appropriateness, and completeness are partly interpretive and cannot be reduced to surface similarity metrics.

At the same time, other studies warn that automated evaluation remains limited. In the course-material-generation framework Instructional Agents, LLM reviewers were “much less discriminative than humans,” with automated evaluators clustering around 2.9–3.1 while human evaluators spread scores more broadly (Yao et al., 27 Aug 2025). The authors therefore use human ratings as the main quality reference (Yao et al., 27 Aug 2025). A plausible implication is that multi-judge LLM evaluation is useful for scaling comparisons, but high-stakes assessment of instructional materials still depends on human review.

Security research adds a different constraint: the same obedience that makes an instructional agent useful can make it exploitable. Work on documentation-following LLM agents identifies the Trusted Executor Dilemma, arguing that high-privilege agents execute documentation-embedded instructions at high rates because they cannot reliably distinguish malicious directives from legitimate setup guidance (Kao et al., 12 Mar 2026). ReadSecBench contains 500 real-world README files across Java, Python, C, C++, and JavaScript, with manually reviewed payloads inserted into setup or requirements sections and Fleiss’ $Rt = (Tr, Ar)$7 for plausibility review (Kao et al., 12 Mar 2026). The paper reports end-to-end exfiltration success rates up to 85%, semantic compliance rates ranging from 46% to 79% across four LLM backends in simulation, 100% successful instruction parsing and attempted execution across four agent systems, and a 0% detection rate in a 15-participant user study (Kao et al., 12 Mar 2026). Evaluation of 12 rule-based and 6 LLM-based defenses did not produce reliable detection without unacceptable false-positive rates. This establishes a direct controversy in the design of instructional agents: strong instruction-following is not unconditionally desirable when the instruction source is only partially trustworthy.

Simulation frameworks for educational validation respond to a related concern. AgentSchool argues that real-world trials of educational AI are slow, ethically constrained, and institutionally locked, and therefore proposes a multi-agent simulator in which learning is modeled as state transition rather than role-play (Ye et al., 28 May 2026). Student agents carry explicit knowledge graphs and misconceptions; teacher agents plan, scaffold, and reflect; scenery shapes the social and instructional field. The broader implication is that future instructional-agent research may need “wind tunnel” environments not only for pedagogical efficacy but also for long-horizon institutional and safety analysis (Ye et al., 28 May 2026).

Taken together, these results indicate that instructional agents should not be understood as a settled application category. The literature instead presents an expanding research domain whose central problems are instructional coherence, learner-state modeling, role specialization, controllability, trustworthy evaluation, and calibrated autonomy. Systems differ widely—from teachable companions and persuasive agents to multi-agent ISD pipelines, embodied professors, classroom instructor agents, and documentation executors—but they converge on a common technical challenge: encoding instructional structure deeply enough to be useful without allowing that structure to become brittle, ungrounded, or unsafe.