Tailoring AI Agents for Early Learning: The Creative Project Approach

Synthesis: Yang, Li, and Lee (2025) offer a conceptual roadmap for tailoring generative AI agents and robotics to early childhood learners, arguing that physical (rather than screen-based) agents are developmentally more appropriate for young children and proposing a "Creative Project Approach" — a five-step pedagogical framework that integrates AI agents into the Project Approach to foster children's creative learning. The paper distinguishes two complementary robotics paradigms rooted in distinct learning theories: coding robots grounded in Papert's constructionism (teaching computational thinking through tangible programming) and generative-AI-powered social robots grounded in Vygotsky's social constructivism (acting as conversational peers or tutors who scaffold cognition and social-emotional development). Teachers remain central as facilitators who scaffold child–robot interaction and preserve child Learner Agency.

Key Findings

1. Physical agents over screens for young children. To mitigate excessive screen time and early screen exposure, Yang, Li, and Lee argue that physical, embodied robots — both coding robots and social robots — are more appropriate than screen-based virtual agents for early childhood education, because they create unique physical- and social-interactive learning experiences.
2. A dual theoretical grounding. The framework is anchored in two learning theories. Constructionism (Papert, LOGO lineage: Bee-Bot, KIBO, Matatalab) underpins coding robots, where children "learn by making" and develop computational thinking — sequencing, loops, conditionals, debugging — through tangible programming. Social constructivism (Vygotsky) underpins generative-AI social robots, which offer Scaffolding within a child's Zone of Proximal Development, acting as a peer or tutor who supports both cognitive and socio-emotional development. Coding robots foreground Problem Solving and STEM; social robots foreground relational interaction and social-emotional growth.
3. The "Creative Project Approach" framework. A five-step model aligned with the Project Approach (Katz & Chard) guides integration of AI agents into early-childhood projects: (1) identify learning needs, (2) facilitate child–robot interaction with teacher guidance, (3) situate AI and robot use in various learning contexts, (4) determine the appropriate level of automation and Creativity, and (5) evaluate learning outcomes. Two illustrative cases show the approach in practice — a story-inspired Matatalab coding project (sequencing and debugging embedded in a picture-book narrative) and a play-based GenAI project where a kindergarten class co-created a birthday song using the Doubao AI voice agent, amplifying Learner Agency and social-emotional growth through collaborative music-making.
4. Roles of generative AI agents. The authors synthesize several agent roles for young children: personalized conversational partners and on-call facilitators that adapt content to each child's pace and emotional state; creative collaborators in co-creating narratives (supporting imagination, language, and decision-making); and social-emotional supporters that simulate social interaction — including robots that scaffold toilet training or support children with autism through safe practice of social skills. Examples span commercial tools (Moxie, BubblePal, Heeyo.ai, Doubao, UBTECH robots).
5. Risks and equity caveats. The paper flags key risks: coding robots' technical fragility and complexity can frustrate young learners and demand teacher training and resources; generative social robots risk hallucination — fabricating content that preoperational children may uncritically accept as truth — and often lack the nuance to calibrate Feedback developmentally, potentially hindering self-regulation and critical thinking. Cost and access concerns widen the Digital Divide and inequity, and data Privacy is a live concern. Teachers need training in both the technological and pedagogical sides, and should infuse cultural relevance so technology aligns with local sociocultural contexts.

What this means for practice

• Instructors. Choose physical robots over screen-based agents for early-childhood AI activities: the framework argues embodied interaction reduces added screen exposure and creates physical and social learning experiences that virtual agents cannot.
• Instructors. Work the five steps in order — identify learning needs, facilitate child–robot interaction with your guidance, situate the use in varied contexts, set the level of automation and creative openness, then evaluate outcomes — instead of starting from a tool you want to use.
• Learning designers. Embed the AI agent inside the project's existing narrative or play frame, as in the two cases presented: coding activities sequenced inside a picture-book story with Matatalab, and a class co-creating a birthday song with the Doubao voice agent.
• Teacher educators. Prepare teachers for both the technical and pedagogical sides of the tools, including how to catch and correct fabrications, because the authors warn that preoperational children may accept hallucinated content as true and that AI feedback is often not calibrated to a child's development.

Limitations

• The paper is conceptual: the framework is derived from theory and prior evidence, and no data were collected on whether its five-step sequence changes children's creativity relative to routine practice.
• The two supporting cases are illustrative accounts — a Matatalab story-based coding project and a kindergarten class co-creating a birthday song with Doubao — reported without participant counts, comparison conditions, or outcome measures.
• The recommended tools are mostly commercial (Moxie, BubblePal, Heeyo.ai, Doubao, UBTECH robots) whose effectiveness evidence is not reviewed here, and the authors themselves note that research on this technology is rarely applied in real-world settings.
• The risk discussion is author-stated rather than measured: hallucination in generative social robots, software and hardware failures in coding robots, and cost and access barriers that widen the Digital Divide are named as concerns but not quantified.

Citation

Yang, W., Li, H., & Lee, J. C.-K. (2025). Tailoring AI agents for early learning: The Creative Project Approach. Computers and Education: Artificial Intelligence, 9, 100473.