Agentic AI Ecosystems in Higher Education: A Perspective on AI Agents to Emerging Inclusive, Agentic Multi-Agent AI Framework for Learning, Teaching and Institutional Intelligence

Synthesis: Sudarshan, Sisodia, Ramachandra, Batra, and Chong Leng Leng (2026) argue that AI in higher education has stalled at the level of the isolated tool. Agents for tutoring, grading, scheduling, and analytics work well inside narrow functions but do not coordinate, so institutions cannot align planning or adaptive decisions across learning, teaching, and administration. They ground this in an inductive thematic analysis of published studies of AI agents in undergraduate education, coded for functional role, architecture, inclusion features, and limitations. Four themes emerge: functional fragmentation of task-specific tools, a shift from single-agent to multi-agent systems, limited cross-functional integration, and thin attention to inclusivity and accessibility. The contribution is a unified multi-stakeholder framework in which learning, teaching, inclusion, and institutional agents form one coordinated ecosystem across a layered architecture of interfaces, agent coordination, and data infrastructure. Its distinctive claim: inclusive practice can be embedded in that architecture rather than added later, so coordinated adaptive and Multimodal AI interventions can support diverse learners, including students with special educational needs.

Key Findings

• Fragmented tools dominate the field. About 52% of studies address task-specific applications such as intelligent tutoring, chatbots, grading systems, and learning analytics, optimizing isolated functions rather than institution-wide intelligence.
• Agentic multi-agent systems are the least studied category. Roughly 28% of the studies concern single agents, about 12% multi-agent systems, and only some 8% agentic multi-agent systems.
• Inclusion is implicit rather than designed. Personalization, accessibility, and affective features appear but remain isolated, and even advanced agentic work rarely addresses cognitive diversity, sensory accessibility, or emotional well-being explicitly.
• Cross-functional integration is largely absent. Individual applications show effectiveness, yet the analysis finds little integration across student-facing learning, educator support, and institutional decision-making.
• Inclusive support is the distinctive contribution. Coordinated agents are argued to enable adaptive, multimodal interventions such as personalized pacing, scaffolding, text-to-speech, and affect-aware well-being support for students with special educational needs.

Four themes from the literature

An inductive thematic analysis coded the reviewed studies for functional role, system architecture, inclusion and learner support features, and limitations. The distribution is uneven: function-specific tools form the largest group at roughly 52%, single-agent systems about 28%, multi-agent systems about 12%, and agentic multi-agent systems only about 8%. The four themes are functional fragmentation, the shift from single-agent to multi-agent architectures, limited integration across learning, teaching, and institutional processes, and thin, inconsistent treatment of inclusive learning. Examples such as GenMentor, MultiTutor, EduVerse, and FACET show that role specialization, memory, reasoning, and coordinated decision-making are technically possible, but most stay at prototype or conceptual stage, with limited large-scale deployment and empirical validation.

A unified multi-stakeholder agentic architecture

The central proposal is one coordinated ecosystem rather than a collection of separate tools. Learning agents customize content to a student's knowledge level, learning style, emotional state, and interests, answer subject questions, and promote Self-Regulated Learning. Teaching agents act as aides to educators, generating lesson plans and quizzes, automating feedback, and surfacing performance insights, and are positioned as assistive rather than replacements. Institutional agents sit at the administrative and strategic level, using predictive analytics to flag at-risk learners, handling enrollment, advising, and routine queries, and optimizing classroom, timetable, and exam scheduling. The defining feature is cross-functional coordination: feedback loops between learning and teaching, alignment of pedagogical practice with institutional objectives, real-time adaptation to learner behavior, and system-level optimization through distributed intelligence. This is a conceptual framework and research agenda, not a deployed platform.

Inclusive learning and special educational needs

Inclusive and equitable learning are treated here as a design problem, not an add-on: students with special educational needs, including learning disabilities, neurodiversity, and sensory impairments, need adaptive environments that current systems rarely provide. Inclusion comes from coordination among specialized agents: accessibility agents adapt interface and modality, cognitive support agents scaffold and pace, well-being agents monitor affect, and learning agents deliver content and assessment. Three dimensions of support follow. Cognitive support covers personalized pacing, simplified explanations, scaffolding, and reinforcement for difficulties with comprehension, memory, and attention. Sensory support covers text-to-speech, speech-to-text, visual enhancements, and alternative content representations. Emotional support covers affect-aware agents that detect stress, disengagement, or overload and respond with feedback, task adjustment, or referral to support resources.

What this means for practice

• Design for coordination, not more tools. Another task-specific agent deepens fragmentation; specify how learning, teaching, and institutional agents share context and hand off work.
• Treat accessibility as architecture. Build multimodal representation, alternative interaction channels, and adjustable pacing into the agent layer, not per request.
• Keep humans in the loop on high-stakes work. The paper flags over-automation in sensitive domains such as mental health as a risk requiring human oversight.
• Plan interoperability and governance together. Heterogeneous platforms, sensitive learner data, bias, and opaque decisions precede deployment.

Limitations

• Perspective, not evaluation. The paper argues a position grounded in thematic analysis of existing literature; it reports no deployed system, and its framework is not tested, validated, or evaluated.
• The synthesis rests on the studies it reviews. The coding covers the studies in the paper's own tables, so the reported distribution reflects that selected corpus, and many cited systems remain prototypes.
• Implementation questions stay open. Interoperability, scalability, pedagogical over-reliance, bias, privacy, and human oversight are unresolved challenges rather than solved problems.

Citation

Vidya K Sudarshan, Anushka Sisodia, Reshma A Ramachandra, Sia Batra, Josephine Chong Leng Leng (2026). Agentic AI Ecosystems in Higher Education: A Perspective on AI Agents to Emerging Inclusive, Agentic Multi-Agent AI Framework for Learning, Teaching and Institutional Intelligence. arXiv:2605.14266. https://doi.org/10.48550/arXiv.2605.14266