Agentic AI in Education: State of the Art and Future Directions

Synthesis: This narrative synthesis (2015–2025) of agentic AI in education provides a definitional and conceptual scaffold for a fragmented field. It operationalizes "agentic" via a six-criterion checklist (autonomy, reasoning, memory, planning, goal-directed action, learning/adaptation) requiring at least four criteria to qualify — excluding reactive chatbots — and organizes the space along three axes: pedagogical role, autonomy level (reactive → adaptive → proactive → collaborative), and embodiment (text-based, avatar/graphical, embodied/robotic). It reviews benefits (personalization, Motivation, teacher assistance) against challenges (over-scaffolding, opacity, bias, equity) and charts future directions toward pedagogically-aware, interoperable, explainable, and ethically governed agents — arguing agentic AI should be a human-AI co-teaching partner, not a teacher replacement.

Core Contribution: An Operational Definition of Agentic AI

A central gap the survey addresses is that "agentic AI" is used loosely. The authors propose a six-criteria operational checklist — a system counts as agentic if it meets at least four: (1) autonomy (action independent of continuous human intervention), (2) reasoning/planning, (3) memory/context-awareness, (4) goal-directed action toward learning outcomes, (5) adaptability to changing learner context, and (6) dynamic collaboration/initiative. The ≥4 threshold is designed to capture the minimum combination of autonomy, reasoning, memory, and action needed for goal-directed educational behavior while excluding reactive chatbots (e.g., a static FAQ bot without planning or persistence would not qualify). A tutoring system that plans multi-turn lessons, remembers learner progress, triggers prompts autonomously, and gives formative feedback would qualify.

A Taxonomy: Role × Autonomy × Embodiment

The survey classifies agentic AI in education along three axes:

• Pedagogical role: tutors (grounded in the ITS legacy but adding proactive scaffolding and adaptive strategy transition); learning coaches/mentors that build metacognitive and affective skills aligned with self-regulated and socio-emotional learning; companion agents for conversational co-construction; instructors' assistants (grading, summarizing, flagging at-risk students, recommending interventions); and curriculum planners/designers that suggest personalized trajectories.
• Autonomy level: a spectrum from reactive (respond only, no memory — rule-based chatbots) → adaptive (learner-model-based short-term feedback loops) → proactive (goal-directed, set sub-goals, replan, provide cross-interaction continuity → dynamic Scaffolding) → collaborative (joint task execution and decision-making with humans or other agents, most common in multi-agent contexts).
• Embodiment/interface modality: text-based Large Language Models (LLMs) agents (platform-independent, low barrier, scalable); avatar/graphical agents that raise social presence and emotional involvement; and embodied agents (classroom robots or AR/VR agents) enabling spatial, gestural, and kinesthetic interaction.

Benefits and Applications

Agentic systems move from task-specific executors to proactive, context-sensitive collaborators. Documented benefits include deeper personalization over extended interactions, adaptive practice, unburdening instructors of routine work (grading, monitoring, recommending interventions), richer Multimodal AI interaction, and scalable Simulation for low-risk practice. The survey notes students respond more favorably to socially present agents when their appearance aligns with pedagogical goals and cultural norms.

Challenges, Risks, and Mitigations

The survey translates pedagogical risks into concrete mitigations with measurable Guardrails:

• Pedagogical misalignment and over-scaffolding: agents giving feedback inconsistent with learning goals, or intervening too frequently, create dependency and reduce Problem Solving and self-AI Regulation in Education. Mitigation: fading protocols that gradually reduce hints (e.g., target a Help-Seeking ratio below 0.3) and regular alignment audits against course objectives.
• Opacity and low trust: opaque 'black-box' reasoning undermines learner and teacher confidence and blocks validation. Mitigation: explainable feedback ("Why this suggestion?"), timestamped decision-traceability logs for instructional auditing, and confidence indicators.
• Bias, fairness, and cultural sensitivity: LLM- and multimodal-based agents can replicate and amplify gender, racial, socioeconomic, linguistic, and cultural bias. The survey emphasizes diverse training data, fairness audits, and cultural grounding.
• Privacy and equity: autonomous agents raise data-AI Governance concerns, and uneven access risks widening digital divides.

Future Directions

The survey charts research and policy directions: (1) pedagogically-aware agents grounded in learning theory (constructivism, cognitive load, self-regulated learning) rather than surface conversation fluency, developed via interdisciplinary AI–learning-science collaboration; (2) interoperable, modular, open platforms with well-documented APIs and standards, to counter siloed proprietary systems that reinforce inequity; (3) multimodal and embodied learning experiences (VR/AR, robots, multimodal feedback and assessment); (4) human-AI co-teaching and hybrid classrooms where agents handle personalized practice and feedback while instructors own classroom dynamics and second-order pedagogical decisions; (5) evaluation and benchmarking frameworks that measure pedagogical quality and longitudinal learning gains, not just language performance; and (6) ethical governance and policy frameworks — transparent data policy, participatory design with students/teachers/parents/administrators, and international cooperation on accountable, auditable deployment. Throughout, the authors argue people-centered values, ethical stewardship, and human-in-the-loop safeguards (instructors monitoring, revising, or overriding agent output) are essential.

Significance to the Knowledge Base

This survey offers the field a much-needed definitional and taxonomic anchor that complements the knowledge base's empirical scoping review: where Wang et al. map the research landscape and capability levels across 474 studies, Kostopoulos et al. supply the conceptual apparatus (operational checklist + role/autonomy/embodiment taxonomy) for classifying and designing agentic systems, and a systematic statement of the design tensions between automation and learning.

What this means for practice

• Designers. Classify a candidate system against the six criteria before calling it agentic: a reactive FAQ bot without planning or persistence fails, and so does an adaptive quiz engine with no memory or goal-directed action.
• Designers. Instrument fading from the start — set target help-seeking ratios (the survey's example is below 0.3) and schedule alignment audits against course objectives — because proactive scaffolding is the documented route from support to learner dependency.
• Instructors. Retain second-order pedagogical decisions — classroom dynamics, assessment weighting, when to override an agent — instead of delegating them, and require explainable feedback and timestamped decision logs from any agent that grades, flags, or recommends.
• Researchers. Evaluate agentic systems on pedagogical quality and longitudinal learning gains rather than language performance, and design against the LLM constraints the survey lists: hallucinated claims on novel topics, limited session-long memory without external memory modules, and predominantly text-optimized multimodal understanding.
• Designers. Favor lightweight, interoperable, open deployments (LTI/xAPI-compatible, open-source or on-device models) over costly cloud-dependent systems, which the survey ties to unequal access across institutions.

Limitations

• The review is narrative rather than systematic: the search covered four databases (Scopus, Web of Science, IEEE Xplore, ACM Digital) from January 2015 to August 2025, and the flow diagram is explicitly "PRISMA-style (simplified)" — no screening counts, no risk-of-bias appraisal, and no pooled effect estimates are reported.
• The ≥4-of-6 criteria threshold is a deliberate definitional cut by the authors, so the taxonomy's boundaries — including the exclusion of reactive chatbots — reflect a design choice rather than a validated construct measure.
• Benefit claims (personalization, motivation, relief of routine instructor work) are drawn from heterogeneous applications the survey neither pools nor weights, so magnitudes cannot be compared across studies.
• The reviewed systems inherit the LLM limitations the authors themselves enumerate: hallucinations when reasoning about new or ambiguous topics, weak session-long memory without external storage, and constrained real-time multimodal understanding.

Citation

Kostopoulos, G., Gkamas, V., Rigou, M., & Kotsiantis, S. (2025). Agentic AI in education: State of the art and future directions. IEEE Access, 13, 177467–177491.