Artificial Intelligence Agents in Computer-Supported Collaborative Learning: A Systematic Literature Review

Synthesis: This systematic literature review of 46 empirical studies (2014–2025) analyzes the roles and impacts of AI agents within computer-supported collaborative learning (CSCL). Guided by the community of inquiry model and learning engagement theory, it finds AI agents most frequently facilitate small-group collaboration and Problem Solving via text-based online platforms, with functions spanning cognitive Scaffolding, social facilitation, and instructional orchestration. While cognitive gains are consistently reported, effects on behavioral, social, and emotional outcomes are context-dependent.

Key Findings

1. 46 studies, post-secondary emphasis. Most research is set in post-secondary settings, with mixed-methods designs most common; AI agents typically facilitate small-group collaboration and problem-solving through text-based online platforms.
2. Three agent functions. AI agents perform cognitive scaffolding, social facilitation, and instructional orchestration, with recent developments enabling more adaptive and participatory roles.
3. Domain-aligned effects. The alignment between agent functions and learning outcomes is strongest within the same domain, though important cross-domain influences are also evident.
4. Cognitive gains consistent, other outcomes context-dependent. Cognitive gains are consistently reported, but behavioral, social, and emotional outcomes appear context-dependent — highlighting the need for nuanced agent design.
5. Need for equity and context-sensitive deployment. The review underscores equitable access, expanded conceptual frameworks, and context-sensitive deployment for meaningful, responsible AI-agent integration in CSCL.

What this means for practice

• Designers. Match the agent's function to the outcome you intend: the review finds cognitive Scaffolding reliably produces cognitive gains and that alignment between agent function and learning outcome is strongest within the same domain, so choose cognitive scaffolding, social facilitation, or instructional orchestration deliberately rather than bundling all three.
• Designers. Treat behavioral, social, and emotional outcomes as context-dependent: cognitive gains were consistently reported across the 46 studies, but the other three outcome classes varied, so pilot the agent in your own setting before assuming social or emotional benefits.
• Designers. Keep groups small: dyads and triads working with a single agent dominate the reviewed evidence, and multiple-agent or larger-group designs remain largely experimental.
• Researchers. Report sample characteristics, agent architecture, and implementation details consistently, and share logging schemas, coding frameworks, and prompting strategies even when raw data cannot be released — almost none of the 46 studies opened their datasets or agent source code.
• Administrators. Treat expansion beyond post-secondary text platforms as governance-intensive: staged validation in K-12, vocational, and multimodal settings needs data minimization, explicit consent, and bias monitoring, and should hold human oversight and equity guardrails as first-order design requirements.

Limitations

• The 46 reviewed studies skew to post-secondary education and text-based online platforms, which the authors state limits immediate generalization to K-12, vocational, and multimodal settings.
• Only 12 of the included studies were large-scale quantitative and 3 were in-depth qualitative; most relied on small institutional datasets, constraining statistical power and generalizability.
• The absence of a control or comparison group was a frequent design constraint in the quantitative and non-randomized studies, limiting causal inference, and participant diversity (cultural background, prior knowledge, demographics) was often not reported.
• Almost none of the included studies provided public access to their datasets, agent source code, or implementation details, limiting reproducibility and comparison across agent architectures.

Citation

Ba, S., Shi, X., Wu, S., & Lu, G. (2026). Artificial intelligence agents in computer-supported collaborative learning: a systematic literature review. Computers and Education: Artificial Intelligence, 10, 100579.