A systematic review of the roles of agentic AI in higher education

Synthesis: This PRISMA systematic review (Jan 2023 – Mar 2026; 48 studies from 1247 records) asks what roles agentic AI actually plays in higher education and how those differ from generative and rule-based AI. It identifies six discrete roles and quantifies the evidence for each: personalized learning and adaptive tutoring (18/48, the most-evidenced), automated assessment and feedback (12), teaching assistance and augmentation (11), administrative and student support (8), curriculum design and workforce alignment (5), and research support and academic operations (4). Reported outcomes cluster on scalable personalization (academic gains 15–25%, engagement up to +40%), assessment efficiency (90–95% human-grader agreement, 50–70% grading-time cuts), and proactive at-risk support — set against risks of cognitive dependency, cultural homogenization, algorithmic bias, and low faculty readiness (only 6.5% of faculty in some contexts report direct AI use). The review's central practical contribution is a tripartite framework for responsible integration: develop AI Literacy, build ethical infrastructure, and deliver competence-based educator training.

A role-based map of agentic AI in higher education

Where other syntheses (the knowledge base's 474-study scoping review and Kostopoulos et al.'s conceptual survey) map the field's breadth, this review organizes evidence by the role an agent plays — a framing tailored to institutions deciding where to deploy and govern these systems. The six roles, in order of evidential weight:

• Personalized learning and adaptive tutoring (18 studies). Agents that continuously model the learner, adjust content/pacing, and give targeted feedback without constant human orchestration. Technical foundations are reinforcement learning, NLP, affective computing, and cognitive modeling. Notable systems include the LMS-embedded ELA Tutor (López-Goyez et al., 2026) and Abertay's tri-modal Learning Engagement Assistant (Chat/Tutor/Quiz). Reviews argue multi-agent cognitive architectures beat single-agent or rule-based tutors on behavior modeling, real-time feedback, and metacognitive enhancement.
• Automated assessment and feedback (12 studies). Rubric-based grading, automated essay scoring, formative feedback, and real-time diagnostics — reported at 90–95% agreement with human graders and 50–70% grading-time reductions, but with persistent concerns about bias and metacognitive erosion from instant, unreflective feedback.
• Teaching assistance and augmentation (11 studies). A spectrum from routine administrative Q&A (virtual TA) to full course delivery by an "AI instructor" — the far end raising questions about educator identity and the limits of automation.
• Administrative and student support (8 studies). Enrollment/scheduling automation plus proactive early-warning agents that monitor data streams and reach at-risk students before problems entrench (e.g., Element451's Bolt ecosystem, Ivy Tech deployments) — inverting traditional reactive support.
• Curriculum design and workforce alignment (5 studies). Agents that scan labor-market data, map course content to competency frameworks, and recommend curriculum change — small but strategically significant.
• Research support and academic operations (4 studies). Literature review automation and multi-agent research frameworks such as Agent Laboratory (Johns Hopkins + AMD), whose coordinated agent team mirrors the research pipeline.

Defining agentic AI: a property-based account

To separate agentic from merely generative or rule-based systems, the review characterizes agentic AI as a class of architectures defined by intersecting properties — complementing the knowledge base's six-criterion checklist with an adjacent framing. The properties: autonomy (acts/decides without per-step instruction), reactivity (perceives and adjusts to its environment), proactivity (initiates from inferred goals, not explicit commands), learning (improves via reinforcement or continual tuning), tool use (invokes search, code executors, schedulers, databases — extending capability beyond the core Large Language Models (LLMs)), memory (short- and long-term, enabling cross-session continuity and personalization over weeks), and planning (decomposing goals into ordered sub-goals). Together these enable what Acharya et al. (2025) call a "qualitative leap" — systems that plan, use tools, hold memory, and self-adjust over multi-step horizons. Notably, the roles-with-framing also surfaces the Copilot-vs-Autopilot design choice: how much moment-to-moment control the human retains versus how much the agent owns — a recurring axis across roles.

Benefits: scalability, personalization, and operational efficiency

Across the 48 studies the reported benefits cluster into three interconnected dimensions. Scalability lets virtual TAs, early-warning systems, and research tools serve many students without proportional staffing — changing the economics of student support. Personalization at scale becomes operationally real: adaptive tutoring reports academic gains of 15–25% and engagement increases of up to +40%, with particular potential for students historically ill-served by one-size-fits-all instruction (first-generation students, learning differences, second-language learners). Operational efficiency appears in 50–70% grading-time reductions and accelerated research workflows. A fourth, cross-cutting opportunity is the shift from reactive to proactive support — intervening on indicators before problems entrench rather than after.

Risks: the passive-learner finding and beyond

The most striking empirical caution is Sargsyan's finding that passive learners in agentic-tutoring environments underperformed Active Learning students by 8.7% — evidence that the risk is not intrinsic to the technology but follows from deployment that lets the agent do the cognitive work. This mirrors the knowledge base's core over-reliance concern and the automation-vs-learning tension. Four risk clusters recur:

• Cognitive dependency and eroded deep learning. Passive consumption of automated feedback without metacognitive engagement — worse outcomes than deliberate active-learning design.
• Cultural homogenisation and epistemic hegemony. Agentic systems trained on English, Western-produced data embed particular assumptions about knowledge, argumentation, and academic register; language-education tools can marginalize non-Western rhetorical traditions, and Global South analyses (Khoza & van der Walt, 2025) show AI-enhanced pedagogies reproducing inequity absent attention to epistemological diversity and infrastructure.
• Algorithmic bias at multiple levels — training data, architecture, evaluation criteria, and test populations — requiring institutional mitigation (bias audits, diverse datasets, transparent documentation, stakeholder involvement) rather than one-time checks.
• Governance lag and data privacy. Agentic tutoring, adaptive assessment, and early-warning require continuous access to granular student data (behavior, performance, engagement, psychological states, even biometrics), yet institutional data-governance frameworks (GDPR/FERPA compliance, Privacy) frequently lag deployment practice.

A tripartite framework for responsible integration

The review's central practical contribution is a synthesis of the evidence into three pillars institutions should stand up before scaling agentic AI — an institutional complement to the design-level Guardrails on the agentic AI concept page:

• Develop AI literacy (students and staff). Literacy is a precondition for meaningful engagement rather than passive consumption: students who understand what an adaptive tutor is doing engage as active learners, and faculty who understand assessment agents design assessments that both exploit and safeguard their capabilities.
• Build ethical infrastructure. Data-protection policies, algorithmic-accountability frameworks, and Academic Integrity policies must precede large-scale deployment. Kasarci et al.'s (2025) four-level governance framework offers a template.
• Deliver competence-based educator training. The professional-development gap is stark — only 6.5% of faculty in some national contexts report direct AI use for education — and training must move beyond tool familiarization to pedagogical frameworks that preserve human agency.

Future directions

The review calls for multi-year cohort studies tracking achievement and self-regulation, and for research on integrated multi-agent campus architectures — orchestrating tutoring, assessment, administrative, curriculum, and research agents within unified systems — since current evidence treats each role in isolation and misses emergent, systemic effects. A plausible five-year scenario: a university running an integrated, human-supervised campus-wide agent ecosystem.

Significance to the Knowledge Base

This review's role-based map and outcome quantification distinguish it from the field's other syntheses: the scoping review maps research breadth and capability levels; Kostopoulos et al. supply the conceptual checklist and taxonomy; and this review adds the deployment-oriented question — which role, with what measured outcome, under what governance — grounding the concept page's design tensions in an institutional, role-specific frame and a concrete responsible-integration agenda.

What this means for practice

• Instructors. Pair any adaptive tutor with tasks that keep students doing the cognitive work: the review's sharpest empirical caution is Sargsyan's finding that passive learners in agentic-tutoring environments underperformed Active Learning students by 8.7%, a gap attributed to deployment rather than the technology itself.
• Administrators. Stand up the review's three pillars before scaling anything: AI Literacy for students and staff, ethical and data-protection infrastructure (GDPR/FERPA-class governance, bias audits, accountability frameworks), and competence-based educator training — the professional-development gap is stark, with only 6.5% of faculty in some national contexts reporting direct AI use for education.
• Administrators. Move student support from reactive to proactive. The eight studies in the administrative-and-support role include early-warning agents that monitor engagement and performance streams and intervene before problems entrench, which changes the staffing economics of at-risk support.
• Designers. Decide the Copilot-versus-Autopilot split explicitly for each role: the review treats the amount of moment-to-moment control retained by humans as a recurring design axis, and the roles where the agent owns the loop (grading, early warning) are the ones where bias and metacognitive erosion concentrate.
• Researchers. Treat the headline outcome figures (15–25% academic gains, up to +40% engagement) as aggregations carried from Yuensook et al.'s synthesis of included reviews rather than as causal effects of agentic systems, and design the multi-year cohort studies the review calls for.

Limitations

• The protocol was set a priori but was not prospectively registered (for example, in PROSPERO), and screening and data extraction were performed by a single author, with a second screener cross-checking only a 15% random sample for reliability.
• The 48 included studies were drawn from 1247 records in five databases (Scopus, Web of Science, ERIC, Google Scholar, Wiley) with a hard March 2026 cut-off; sources found in the July 2026 search update were added narratively rather than re-entered into the screening flow, so the pathway counts describe the earlier corpus only.
• The corpus is geographically skewed — North America 35%, East and South Asia 28%, Europe 21%, Global South 10% — a concentration the authors name as a notable limitation, and adaptive-tutoring evidence is dominated by STEM and language learning, leaving the humanities, social sciences, and professional education under-evidenced.
• Agentic status was assigned by a threshold rule rather than a validated taxonomy — a study counted as agentic only if its system met at least three of four criteria (autonomy, planning, tool use, persistent memory) — so the six role counts depend on that binary classification judgment, and vendor documentation and technology reports are cited illustratively in the Results and Discussion without being counted among the 48 included studies. The evidence base the review synthesizes is itself methodologically young — mostly short-term case studies and design demonstrations, few RCTs and a near-total absence of longitudinal work — so causal claims about learning outcomes cannot yet be made.

Citation

Baradziej, S. (2026). A systematic review of the roles of agentic AI in higher education. Social Sciences & Humanities Open, 14, 103418.