Agentic AI in higher education: Transforming interactive learning while navigating implementation challenges

The next phase of artificial intelligence in higher education is not simply about generating content in response to prompts. Agentic AI systems can pursue goals, plan actions, retain context, adapt to changing learner needs, and act with a degree of autonomy. This shift moves AI beyond a reactive tool towards a potentially active participant in the learning process. Its significance lies not in autonomy alone, but in how institutions configure that autonomy to support sound pedagogy, learner agency, and responsible decision-making. 

In education, Agentic AI can move from passive support to a co-creative learning partner. It may provide direct learner support as a tutor, mentor or simulated peer; assist educators with teaching and curriculum design, and support institutional activities such as identifying disengagement or recommending targeted intervention. This article argues that its educational value depends on three interconnected design dimensions: pedagogical role - what the agent does; level of autonomy - how independently it can act; and embodiment and interaction - how it engages with learners. Together, these dimensions provide a practical framework for pedagogy-led adoption across the learner journey. 

Three paradigms for Agentic AI in learning and teaching 

1. Pedagogical role: What function does the agent perform? 

Agentic AI delivers instructional content, guides learners through structured tasks and provides timely formative feedback. These systems draw on the existing intelligent tutoring systems but differ with proactive scaffolding and adaptive strategy transitioning in response to context. Agentic AI can deliver instructional content, guide learners through structured tasks and provide timely formative feedback. Building on intelligent tutoring systems, it can add proactive scaffolding, flexible instructional strategies and responsiveness to context. Used as a learning coach or mentor across programmes and modules, an agent may help learners develop metacognitive and affective capabilities. It can support a dyadic learning relationship in which learners set goals, monitor progress and persist through productive struggle, consistent with principles of self-regulated and socio-emotional learning. 

In practice, this dimension may include: 

• Learning companions and simulated peers: Agents can participate in collaborative scenarios, model effective learning practices and support social-emotional learning.

• Teaching and classroom assistants: Agents can help educators design more personalised interventions, monitor participation and direct attention towards learners who may require additional support. 

• Curriculum planning agents: Agents can analyse learner data, recommend personalised learning pathways and adjust instructional sequences in response to learner progress and classroom dynamics. 

2. Level of autonomy: How independently can the agent act? 

Proactive behaviour and autonomy are related, but they are not identical. An agent may anticipate a learner’s needs while remaining subject to educator approval, or it may be authorised to set sub-goals, replan actions and adapt support within defined boundaries. A responsible model therefore distinguishes between recommendation, guided action with human approval, bounded autonomous action and high-stakes decisions that require human oversight. The aim is not to maximise autonomy, but to calibrate it to the pedagogical purpose, the consequences of error and the learner’s capacity for independent judgement. 

In practice, bounded autonomy may support: 

• Continuity across the learner journey: Within agreed limits, agents can track skill development over time, revisit learning objectives and recommend pathways aligned with academic goals and emerging professional aspirations. 

• Simulation-based learning and virtual laboratories: Agents can support immersive, interactive environments in which learners apply knowledge, practise professional judgement and receive adaptive feedback, while digital portfolios record developing capabilities. 

3. Embodiment and interaction: How does the agent engage with learners? 

Agentic AI can engage through text, voice, images and gesture, creating multimodal interactions that may improve accessibility and participation for learners with diverse needs and preferences. Three forms should be distinguished: multimodal agents, which communicate across text, voice or visual interfaces; immersive agents, which operate in augmented or virtual reality; and physically embodied agents, such as classroom robots. Each form offers different educational affordances, costs and data implications. Used purposefully, these agents may support dialogue, spatial interaction, kinaesthetic learning and more natural communication; however, the collection of granular behavioural or biometric data requires proportionate governance. 

In practice, this dimension may include: 

• Dialogue-based learning: Multimodal agents can support Socratic questioning, adopt an appropriate simulated persona and facilitate collaborative problem-solving within an active, learner-centred environment. 

• Immersive learning: Agents in augmented or virtual environments may create a stronger sense of presence and provide context-sensitive feedback, particularly in simulations that require rehearsal, reflection, or spatial interaction. 

No educational technology is without risk. The institutional challenge is therefore not simply whether to adopt Agentic AI, but how to align its role, autonomy and embodiment with educational purpose. Effective implementation requires clear governance, proportionate human oversight and continuing evaluation. 

Implementation challenges and governance priorities 

• Pedagogical alignment: Feedback may be inaccurate, vague or disconnected from learning outcomes. Institutions should validate agents against curriculum requirements, approved knowledge sources and assessment criteria, with educators retaining oversight of consequential guidance. 

• Learner agency and dependency: Excessive scaffolding may reduce productive struggle, critical thinking and self-regulation. Agent support should be deliberately faded, with learners required to explain decisions, evaluate advice and retain responsibility for their work. 

• Accountability and human oversight: Autonomous recommendations can become incoherent or inappropriate when separated from the wider instructional context. Institutions should define decision rights, escalation routes and approval thresholds, particularly where an agent contributes to intervention or assessment. 

• Privacy and data stewardship: Personalisation may rely on interaction logs, performance data, behavioural cues or biometric information. Adoption should follow clear consent, data-minimisation, retention, access and security regulations.  

• Bias, fairness and accessibility: Agent recommendations and interfaces may perform unevenly across learners. Regular bias testing, accessibility review and routes for challenge are essential to ensure that support does not reproduce or amplify disadvantage. 

• Institutional readiness: Effective adoption depends on staff capability, system integration, procurement standards and sustained evaluation. Staged pilots, AI literacy and transparent success measures should precede institution-wide deployment.