Traditional support systems often rely on decision trees or single-model architectures that struggle when query complexity exceeds simple keyword matching. Fanatics Betting and Gaming (FBG) encountered this limitation as their user base expanded across multiple U.S. states, each with distinct payment rules, deposit limits, and responsible gaming requirements. To address the surge in inquiries during high-traffic events like NFL playoffs without degrading response quality or increasing operational costs linearly, FBG engineered a multi-agent system on AWS.
Decoupling Agent Logic via Orchestration
The core architectural change involves moving away from monolithic chatbots toward an orchestrator pattern. In this design, a primary Supervisor Agent coordinates with specialized sub-agents and tools rather than attempting to handle every query type within one model instance.
FBG leveraged their existing container platform on Amazon Elastic Kubernetes Service (Amazon EKS) for deployment independence. By running agents independently, the team can iterate on specific capabilities—such as a new knowledge domain or business unit tooling—without rewriting core system logic. This modularity is critical when scaling support volume; adding a new case type becomes an additive task rather than a systemic refactor.
Guardrails and Compliance Classification
A significant operational implication of this architecture involves the integration of safety layers before model invocation. The request flow routes through Amazon Bedrock Guardrails to detect prompt injection attempts, ensuring that user inputs do not compromise system integrity or data access controls.
Beyond security filtering, a Responsible Gaming classification agent powered by Amazon Nova 2 Lite evaluates every message against an approved compliance framework. High-severity classifications trigger immediate transfers to human agents with full conversation context preserved. This separation of concerns ensures that safety and regulatory logic are distinct from the conversational generation layer.
Model Selection via Bedrock
The system utilizes Amazon Bedrock for its model-agnostic access, allowing FBG to match specific tasks—such as intent classification or retrieval augmented generation (RAG)—to the most appropriate foundation models. This flexibility supports swapping underlying models without altering application code.
What This Means For Practitioners
The move toward multi-agent orchestration offers a blueprint for teams facing similar scaling challenges in regulated industries. By isolating agent responsibilities and utilizing an orchestrator pattern, platform engineers can manage complexity more effectively than with monolithic approaches. Security practitioners should note that while guardrails provide essential filtering against prompt injection, they complement rather than replace downstream authorization controls.
For those evaluating this architecture on AWS, the integration of MCP servers for account data and RAG pipelines demonstrates how retrieval mechanisms must be tightly coupled to agent intent without conflating them with identity boundaries. This approach supports rapid iteration while maintaining strict governance over sensitive customer interactions.
