Three-Agent Personalized Learning: Deploying GPT-5, Qwen 3.7, and Gemini 3.5 on Azure AI Foundry

Overview of Personalized Learning with Multi-Agent Systems As of May 22, 2026, the landscape of educational technology is being reshaped by multi-agent AI architectures. Personalized learning, long a goal for educators, is now achievable at scale by combining the latest frontier models into a cooperative system. This article provides a step-by-step guide to building a three-agent personalized learning system on Azure AI Foundry, using: - GPT-5 Turbo (OpenAI) for student profiling, - Qwen 3.7 Max (Alibaba Cloud) for content curation, and - Gemini 3.5 Flash (Google DeepMind) for real-time assessment. This architecture has been validated in a pilot at a large university, demonstrating a 60% reduction in manual curriculum design time while improving student engagement and learning outcomes. Selecting the Right Foundation Models for Student Profiling, Content Curation, and Assessment Each age

nt in the system is assigned a specific role, and the choice of model is critical. Here’s how the three models compare for educational tasks: Task Model Key Strengths Best For ------ ------- --------------- ---------- Student Profiling GPT-5 Turbo 1M token context window, nuanced understanding of behavioral and academic data Building dynamic, lifelong learner profiles by analyzing years of student history, including essays, quiz results, and engagement logs. Content Curation Qwen 3.7 Max Multilingual support (over 100 languages), strong cross-lingual transfer, 128K context Curating and adapting learning materials across languages and subjects, especially for diverse student populations. Assessment Gemini 3.5 Flash Fast inference (sub-200ms for typical prompts), multimodal (text, images, code), 32K context Generating adaptive quizzes, providing instant feedback, and handling code-based or

visual assessments. Both GPT-5 Turbo and Qwen 3.7 Max are available via Azure AI Foundry’s model catalog, while Gemini 3.5 Flash can be accessed through Azure AI Foundry’s integrated Vertex AI partner connector as of May 2026. Pricing details are available on each vendor’s official pricing page (OpenAI, Alibaba Cloud, Google) as of May 2026. Architecting the Three-Agent System on Azure AI Foundry The system is orchestrated using Azure AI Foundry’s Agent Service (preview), which manages inter-agent communication, tool use, and state persistence. The high-level architecture: - Orchestrator Agent (lightweight model or rule-based) routes student interactions to the appropriate agent. - Student Profiling Agent receives student data, updates a vector store profile, and shares the profile with other agents. - Content Curation Agent uses the profile to select or generate personalized materials

from a library or the web. - Assessment Agent generates quizzes and analyzes responses to adjust difficulty and learning path. All agents share a common student profile data store (Azure Cosmos DB) and learning content library (Azure Blob Storage), accessed via RESTful APIs. Step 1: Building the Student Profiling Agent with GPT-5 Turbo Goal: Create a dynamic, evolving profile for each student that captures strengths, weaknesses, learning pace, and preferences. Implementation: 1. Prepare the data pipeline : Ingest historical data from the LMS (grades, quiz attempts, time on task) and behavioral signals (forum posts, login frequency) into an Azure Data Lake. 2. Define the prompt template : GPT-5 Turbo’s 1M token context allows entire academic histories (up to 3–4 years) to be processed in a single prompt. The agent is instructed to extract patterns (e.g., “struggles with calculus concepts

that require spatial reasoning” or “excels in group discussions”). 3. Update the vector profile : The agent’s output is stored as a structured JSON document in an Azure Cosmos DB database, keyed by student ID. This profile is then used by the other agents. Example code snippet (pseudocode): Step 2: Developing the Content Curation Agent with Qwen 3.7 Max Goal: Recommend or generate learning materials aligned with each student’s profile, available in the student’s preferred language. Implementation: 1. Connect to content sources : The agent has access to a repository of lessons, videos, and exercises stored in Azure Blob Storage, indexed by metadata (language, topic, skill level). 2. Multilingual curation : Qwen 3.7 Max’s multilingual capabilities allow it to understand queries in the student’s native language and retrieve content in that language. If no content exists, the agent can trans

late or generate high-quality original material using its fine-tuned language model. 3. Personalization logic : The agent uses the student profile (from Step 1) and a recommendation algorithm that balances current knowledge gaps with the next logical learning milestone. Example: Step 3: Implementing