ZEP Enhances Educational Quiz Generation Architecture with LangGraph on Amazon Bedrock AgentCore

The metaverse platform ZEP transitioned its architecture from a traditional single-prompt method to a LangGraph-based workflow to improve the quality of ZEP QUIZ, its educational quiz generation service. This shift aims to structurally manage educational constraints such as reflecting misconceptions in incorrect answers, adjusting difficulty levels, and aligning with achievement standards, rather than relying on simple generation.

Separation of Logic and Execution

The system is clearly divided into a LangGraph layer responsible for domain logic and an Amazon Bedrock AgentCore Runtime layer responsible for infrastructure. LangGraph explicitly controls the processes of question generation, verification, and revision through nodes and edges, while AgentCore Runtime executes these as containers to manage authentication, isolation, and scaling.

Multi-Mode Execution

While maintaining a single graph logic, the system simultaneously operates two execution modes.

• Thorough mode: Calls Claude Sonnet 4.6 from Amazon Bedrock to prioritize complex educational context interpretation and stable structured output.
• Fast mode: Directly calls third-party models within the FastAPI process to ensure short response times.

Through this structure, ZEP integrates the management of real-time user requests and background batch generation for new content into a single logic, structurally verifying the uniqueness of correct answers and the consistency of evidence for generated questions.