GamED.AI: A Hierarchical Multi-Agent Framework for Automated Educational Game Generation

Abstract: We introduce GamED.AI, a hierarchical multi-agent framework that transforms instructor-provided questions into fully playable, pedagogically grounded educational games validated through formal mechanic contracts. Built on phase-based LangGraph sub-graphs, deterministic Quality Gates, and structured Pydantic schemas, GamED.AI supports two template families encompassing 15 interaction mechanics across spatial reasoning, procedural execution, and higher-order Bloom's Taxonomy objectives. Evaluated on 200 questions spanning five subject domains, the system achieves a 90% validation pass rate, 98.3% schema compliance, and 73% token reduction over ReAct agents (${\sim}$73,500 $\rightarrow$ ${\sim}$19,900 tokens/game) at $0.46 per game. Within this model configuration, these results suggest that phase-bounded architectural structure correlates more strongly with alignment quality than prompting strategy alone. Our demonstration lets attendees generate Bloom's-aligned games from natural language in under 60 seconds, inspect Quality Gate outputs at each pipeline phase, and browse a curated library of 50 games spanning all 15 mechanic types.

• The paper introduces a hierarchical multi-agent framework that structurally guarantees Bloom’s-aligned educational game generation.
• It employs deterministic Quality Gates and Pydantic schemas to validate game blueprints and enforce mechanic contracts, achieving a 73% token cost reduction.
• The system demonstrates robust performance with a 90% validation pass rate and scalable dynamic model assignment, though real-world classroom efficacy remains to be tested.

GamED.AI: Hierarchical Multi-Agent Framework for Automated Educational Game Generation

Architectural Rationale and Framework Overview

GamED.AI addresses substantive deficits in automated educational game generation: previous general-purpose agentic systems are structurally incapable of guaranteeing pedagogical alignment, primarily due to absent Bloom’s Taxonomy targeting, lack of enforceable mechanic contracts, and brittle self-correction loops. The hierarchical framework designed with phase-bounded LangGraph DAGs, deterministic Quality Gates, and rigorous Pydantic schemas establishes a structurally enforceable pathway from instructor intent to validated interactive game instances. By separating game design, blueprint validation, scene content generation, asset synthesis, and assembly into isolated phases, GamED.AI eliminates error propagation endemic to sequential and reagent architectures.

The system operates model-agnostically, enabling dynamic per-agent model assignment and open-source/closed-source interchangeability without architectural modification. This modularity underpins rapid template expansion and robust game library curation. Games are generated from natural language prompts or learning objectives, grounded via contextual retrieval in domain ontologies, and validated for contract compliance using FOL-based predicates. Inter-agent communication adheres to typed schema boundaries, with parallel dispatch patterns for scalable scene and asset creation.

Template Families, Mechanic Contracts, and Bloom’s Alignment

GamED.AI’s generative surface spans two template families: Interactive Diagram Games (10 mechanics) and Interactive Algorithm Games (5 mechanics). Each mechanic is mapped to a validated set of Bloom’s objectives, ensuring interaction types constitute evidence for the intended cognitive skill. Mechanic contracts are produced prior to content generation, encoded in Pydantic blueprints that specify required primitives, data types, Bloom’s level, and completion triggers. The contract enforcement is structural, not heuristic: only games passing deterministic Quality Gates can propagate to next phases. This tight coupling between design intent and syntactic validation enforces pedagogical primacy and prevents semantic drift.

Template selection and mechanic composition are resolved contextually, supporting single/multi-scene and single/multi-mechanic configurations. Multiscene games comply with cognitive load bounds (≤4 scenes, ≤3 mechanics/scene), and complex scenarios guarantee monotonic Bloom’s level progression. The modular game engine registers each mechanic as a self-contained React component, facilitating extension and enforcing accessibility via WCAG standards.

Evaluation: Structural Guarantees, Token Efficiency, and Human Judgement

A 200-question corpus spanning biology, history, CS, mathematics, and linguistics was used to benchmark GamED.AI against manual authoring, commercial platforms, template-based AI systems, and internal sequential/ReAct baselines. The system achieved:

• Validation pass rate (VPR): 90.0%, a 17.5pp gain over ReAct Agents (72.5%) and 33.3pp over Sequential Pipelines (56.7%), statistically significant ($\chi^2(2, N=600) = 57.0, p < 0.001$, Cramér's $V = 0.31$). This metric exclusively reflects FOL-based structural validators—not pedagogical efficacy.
• Token and cost efficiency: 73% reduction relative to ReAct (73,500→19,900 tokens/game, $\eta^2 = 0.87$), with per-game cost at $0.46—substantially below both manual production ($50–150/unit$$) and platform subscriptions.</li> <li><strong>Schema compliance:</strong> 98.3<li><strong>Human ratings:</strong> 4.2/5 mean educational correctness and playability, statistically similar to manual authoring (4.3/5,$$t(199) = 1.04$,$p = 0.30$); expert rater ICC > 0.78.

Mechanic-specific VPR varied (diagram: 96.2% for drag&drop, 60.0% desc-matching; algorithm: 94.4% state tracer, 80.0% constraint puzzle), with schema underspecification—not model hallucination—as dominant failure mode.

Against Claude Code (with/without mechanic contracts), GamED.AI exhibited a 67pp gap in zero-shot Bloom’s alignment (90% vs 23%) at lower cost and token use, demonstrating that structural validation and bounded generation phases confer guarantees that prompting strategies alone cannot deliver.

Analysis: Implications and Future Directions

GamED.AI demonstrates that hierarchical, contract-driven multi-agent orchestration yields architectural guarantees in educational game generation that are unattainable with legacy prompt-based or unstructured agentic paradigms. The correlation between phase-bounded design and Bloom’s compliance substantiates the premise that formal validation and modularity are prerequisites for pedagogically aligned generative systems.

Practically, GamED.AI compresses hours of expert labor and substantial financial overhead into sub-minute, sub-dollar dynamic game authoring. The modular engine and open-source release facilitate reproducibility, extensibility, and model substitution—directly lowering cost and expertise barriers for EdTech adoption. However, current evidence, while robust in architectural metrics, is limited regarding learning outcome measurement; pedagogical effectiveness in real classrooms remains unproven pending future classroom trials.

Theoretical implications include the realization that formal logic-based validation and deterministic gating are necessary for agentic content-generation architectures in education; future research may extend these principles to domains demanding outcome-linked evidence protocols, such as assessment or simulation-heavy content. The planned expansion to human-in-the-loop negotiation, physics engines, additional mechanic families, and multilingual capability will further generalize the framework and its applicability.

Conclusion

GamED.AI introduces a phase-bounded hierarchical multi-agent framework that structurally guarantees Bloom’s-aligned educational game generation, validated through deterministic mechanic contracts and FOL-based Quality Gates. The architecture attains significant gains in validation, token efficiency, and unit cost versus both agentic and manual baselines, without sacrificing expert-assessed educational correctness or playability. As open-sourced, model-agnostic infrastructure, the system represents a validated authoring tool. Future work will address schema extension, classroom efficacy, and extended template scope, enabling rigorous outcome-linked evaluation and broader real-world deployment.