OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education

Designing effective GenAI agents requires marrying AI capabilities with established pedagogical principles and interaction design frameworks. In this section, we outline key design principles for LLM-powered educational agents, including transparency of AI decisions, scaffolded questioning techniques, multimodal engagement, and maintenance of learner agency. We then examine how these principles are implemented in practice, referencing frameworks like ARCHED and the structured prompt templates (often JSON-based) used in the Socratic Playground system (see Box 3.1). By emphasising features such as conversational pacing, synchrony between verbal and non-verbal cues, and interactive learner controls, we showcase how GenAI agents can balance open-ended dialogue with instructional rigor, thereby fostering learner trust and autonomy even while the interaction is AI-driven.

The ARCHED and other similar frameworks ARCHED (AI for Responsible, Collaborative, Human-centred Education Instructional Design) is a framework to counter the “black box” opacity of automated educational tools by establishing a transparent, human-in-the- loop workflow grounded in Bloom’s taxonomy. Rather than fully automating course creation, the framework employs a dual-agent architecture – comprising a Learning Objective Generation System for drafting content and an Object Analysis Engine for pedagogical auditing – to ensure that educators preserve the autonomy to curate and refine AI-generated objectives. Likewise, other notable frameworks in the field include the efficiency-focused GAIDE (Generative AI for Instructional Development and Education) model and the quality-focused Academic-GPT collaboration framework (adapting the design-build-test-learn approach).

JSON prompt Text prompts are flexible, conversational, and human-friendly but they can leave room for ambiguity. In contract, JSON (JavaScript Object Notation) prompts use explicit fields and machine-readable syntax, which reduces guesswork for the GenAI agent and makes outputs more structured and reliable.

Transparency and explainability

As AI tutors become more complex, transparency in their operation is crucial for building trust with both learners and educators. The ARCHED framework proposes a human-centred approach that embeds transparency and human oversight into AI-assisted instructional design (Li et al., 2025) (see Box 3.1). Within the framework, multiple specialised AI components recommend pedagogical actions and evaluate them, while human educators remain the ultimate decision-makers, ensuring the reasoning behind AI-generated content is visible and can be vetted.

Translating this to a digital tutoring agent scenario, a GenAI agent should ideally be able to explain why a particular question is asked or why certain feedback is given – or at least do so if the learner inquires. For example, an agent might preface a hint by explaining that it is intended to support clarification of the learner’s understanding of a specific concept. Such meta-dialogue provides insight into the agent’s pedagogical intent. Another aspect of transparency is indicating uncertainty. If the AI is not fully confident in a response (which can be estimated from model probabilities or a validation step), it can disclose that uncertainty – e.g. “Let’s double-check this answer, as I’m not entirely sure”. This honesty can help set the right expectations and invite joint problem-solving, rather than the student taking every AI statement as gospel.

Several modern systems have introduced mechanisms to promote their transparency and explainability (See Box 3.2).

• Khanmigo (from the United States): Aiming to provide safe, institutional-grade tutoring, Khanmigo enhances transparency by providing inline citations to external sources and using a “double-check” mechanism where the AI visibly verifies its own answers before displaying them; • Squirrel AI (from China): Intending to optimise learning efficiency through granular diagnosis; it utilises “Explainable AI” to show teachers the reasoning behind specific content recommendations, though the underlying “Large Adaptive Model” generally operates as an opaque system for the student; • Riiid (from South Korea): Focused on efficiently maximising standardised test scores, Riiid employs “Explainable AI” techniques to visualise for users exactly which knowledge components or study behaviours contributed to their predicted scores; • Century Tech (from the United Kingdom): Intended to augment teacher capabilities and personalise learning, their platform ensures transparency by explicitly distinguishing between teacher-assigned tasks and AI-recommended “nuggets” on the student dashboard, identifying the specific skill gaps from “micro-level” learning behaviours that triggered the recommendation.

Alternative to the mechanisms used in existing tools, the implementation of generative pedagogical agents can incorporate mechanisms for post-hoc validation of interactions. In designing SPL, the developers introduced a logging and visualisation tool for researchers and instructors that showed the agent’s decision path (e.g. which prompt pattern was triggered; what the agent “thought” the student’s misconception was). While this backend transparency is not directly available to learners, it allows continuous human oversight of the AI’s pedagogical actions. Overall, adopting a transparent design entails making both the system’s internal reasoning and its external interactions as interpretable as possible, aligning with calls for trustworthy AI in education (Khosravi et al., 2022; Memarian and Doleck, 2023; OECD-Education International, 2023).

Scaffolded questioning techniques

A cornerstone of AI-powered pedagogical agent design is the use of scaffolded dialogue, often drawing from Socratic questioning and related strategies. Rather than delivering answers outright, a well-designed AI-powered pedagogical agent guides the learner to construct knowledge through carefully sequenced questions. This approach is rooted in Vygotskian scaffolding and the Zone of Proximal Development, where support is provided just beyond the learner’s current ability and gradually withdrawn as competence grows (Vygotsky, 1978). LLM- based agents are particularly well-suited to implementing Socratic questioning, as they can generate an extensive range of probing questions and follow-ups dynamically. They can also flexibly rephrase or adjust the difficulty of questions based on learner responses.

Frameworks for intelligent tutoring often include taxonomies of questions (e.g. conceptual probes, evidence requests, counterfactual prompts) that can be encoded into the AI’s prompt or decision logic. In practice, SPL uses a JSON-based prompt template to enforce a structured tutoring script while still leveraging generative flexibility (Hu, Xu and Graesser, 2025). The prompt is divided into sections, such as “Initial_Interaction”, “Following_Up”, “Providing_Feedback”, etc., each containing guidance for the type of Socratic moves the agent should make. For example, in a “Following_Up” turn, the agent might be instructed (via the prompt) to ask a “why” question related to the last student statement, or to request clarification if the student’s answer was incomplete. By structuring the interaction in this way, the agent’s generative outputs remain pedagogically purposeful. More importantly, the JSON structure also allows the system to track expectations and misconceptions explicitly, that is the agent keeps lists of the key points (“expectations”) the student should mention in an ideal answer as well as known common errors (“misconceptions”). Each student response is compared (via the LLM or supplementary classifiers) against these lists, and the subsequent prompt is generated accordingly – e.g. if a misconception is detected, the following question might target that misunderstanding. This method, inspired by AutoTutor’s expectation-misconception tailoring (Graesser et al., 2005) but modernised with LLM capabilities, ensures the question scaffolding is adaptive to the learner’s input. Empirical studies have long shown the effectiveness of such scaffolding as it keeps the learner in an active constructive mode rather than a passive one, which is known to enhance learning outcomes (Chi, 2009).

Adopting the scaffolding approach in agent design aligns with a broader body of research that aims to leverage LLM- driven agents to foster deeper understanding and self-directed learning (Córdova-Esparza, 2025). In designing a GenAI agent, educational technology developers should thus curate a bank of pedagogically sound questioning strategies and incorporate them either through prompt patterns, few-shot exemplars, or rule-based overlays on the LLM’s output.

Multimodal and embodied interaction

To truly advance beyond single-modality interactions enabled by traditional AI tutors, GenAI agents can leverage multimodal engagement – combining text or speech with other modalities like visuals, gestures, or interactive simulations. Research in multimedia learning has shown that well-coordinated verbal and visual information can enhance understanding, as long as they are synchronous and not overwhelming (Mayer, 2002). Modern AI platforms allow a tutoring agent to display images, diagrams, or even manipulate virtual objects in a simulated environment alongside the dialogue. For example, if a student is learning geometry, the agent might dynamically generate a diagram of a triangle and mark angles as it guides the student through a proof. Generative models can produce descriptions of visuals or request relevant images (via integration with image search or generation models), effectively acting as a bridge between text and visuals. Furthermore, if the agent is instantiated as a virtual tutor – whether through AR/VR or screen-based interfaces – the alignment of facial expressions and gestures with dialogue constitutes an important factor in achieving natural interaction. A nod or encouraging smile rendered on the AI tutor’s avatar can reinforce the tone of the agent’s message (e.g. affirming the learner’s progress). Yet, it is worth emphasising that the timing of these cues should align with the conversational content to avoid cognitive dissonance.

The Socratic Playground’s current implementation is primarily text-based with a simple animated avatar representing the AI pedagogical agent, but the design guidelines call for gesture-text synchrony in future versions – for instance, having the avatar produce a “thinking” expression when posing a difficult question, or a cheerful expression when giving positive feedback. The literature on embodied conversational agents (e.g. Krämer, Simons and Kopp, 2007) suggests that such non-verbal behaviours, when congruent with the dialogue, can increase learner engagement and trust in the agent. Nonverbal behaviours are used in several innovative platforms: the DALverse project establishes an inclusive metaverse environment for distance education, where students can engage as digital avatars in multimodal learning tasks, leading to increased engagement and retention in distance learning settings (Damasceno et al., 2024).

The design implications are clear: GenAI agents should, whenever possible, be integrated into interfaces that leverage multiple modalities (e.g. text, voice, graphics) to enable richer learning interactions. However, designers must adhere to established principles of multimedia learning to ensure that these modalities complement rather than compete with one another – for example, by avoiding extraneous animations or redundant narration that merely reads on- screen text aloud, both of which can contribute to cognitive overload.

Maintaining learner agency

A frequent criticism of AI tutors is the risk of learner passivity – if the agent does too much, students might become disengaged or over-reliant on the AI. Therefore, a central design principle is the preservation and promotion of learner agency. GenAI agents can support this in several ways. One approach is through developing learners’ metacognitive awareness. This may be achieved by, for example, posing open-ended questions that allow learners to guide the direction of the interaction, thereby fostering their awareness to steer their own learning journey. Even simple prompts, such as “Would you like another hint or should we try a different problem?”, position learners in an active decision-making role. Interfaces can further enhance agency through interactive controls. For instance, the SPL’s interface offers learners with options to request a simpler explanation, pose a question to the agent, or indicate that they wish to attempt the solution independently. These controls act as safety valves so the student can modulate the help level. Under the hood, the agent monitors these inputs and adjusts its strategy – if a student repeatedly asks for simpler explanations, the agent will reduce the complexity of its language or break problems into smaller steps; if the student wants to proceed independently, the agent will step back and take on a more observational role, ready to jump in only if asked.

Another technique to maintain agency is via implementing turn-taking policies that ensure the AI does not dominate the dialogue. For instance, after the agent poses a question, it should give the learner ample time to think and respond, rather than immediately filling silence with more talk. If a student seems stuck, the agent may offer a hint, but ideally after encouraging the student to articulate any partial thinking first. This aligns with the AI tutoring technique of offering minimal help to keep the student doing as much cognitive work as possible: the goal of such systems is to reach an “interactive” level of engagement, where the student and tutor co-construct knowledge (Chi, 2009). From a design perspective, such an implementation can be useful to measure the proportion of conversation led by the student vs. the agent; some prototype evaluations of SPL looked at what percentage of words or turns were student-generated and aimed to maximise that over time through interface tweaks. Furthermore, the agent can foster agency by being explicitly reflective: encouraging students to set goals, ask their own questions, or evaluate the agent’s suggestions. For example, the agent might say, “Do you agree with the approach I just suggested, or do you think there’s a better way?” – prompting the learner to critically assess the AI’s previous responses, thereby treating the student as an active participant with agency, not just a recipient of knowledge.

Summary

The design of generative pedagogical agents must carefully blend AI innovation with human-centric educational principles. Frameworks like ARCHED provide macro-level guidance on maintaining transparency and human control in AI-powered education systems (Li et al., 2025). The micro-level design in agentic systems such as SPL (Hu, Xu and Graesser, 2025) illustrates concrete features that enact those principles (for example, structured prompts, adaptive questioning, interface controls). By prioritising explainability, scaffolded interaction, multimodal engagement, and learner agency, designers can create AI tutors that are not only powerful and adaptive, but also pedagogically sound and user-friendly. As subsequent sections will show, these design considerations play a critical role in addressing the challenges and ethical implications of generative AI tutors, ensuring that technology serves as a complement to effective teaching rather than a detour from it.