How Fanatics Betting and Gaming constructed a multi-agent buyer help system

Fanatics Betting and Gaming (FBG) constructed a multi-agent buyer help system on AWS to resolve a problem distinctive to sports activities betting. Customers count on on the spot, correct solutions, particularly throughout stay occasions when each minute counts. Customers ask about account points, deposit limits, state-specific laws, and accountable gaming assets. The guidelines range throughout each jurisdiction the place an operator is licensed. Traditional chatbot options constructed on determination bushes battle with this complexity, typically irritating prospects and driving up prices as human agent queues develop.

Fanatics Betting and Gaming (FBG) is a sports activities betting platform that mixes superior expertise with deep sports activities experience. As a part of the Fanatics household of manufacturers, FBG operates throughout a number of U.S. states, serving a quickly rising person base that calls for across the clock help, notably throughout high-traffic occasions like NFL playoffs and the Super Bowl.

Facing exponential development in help quantity, FBG’s engineering workforce constructed a multi-agent AI system on AWS that resolves buyer points sooner, extra precisely, and at a fraction of the price of human-only help. In this put up, we stroll by way of the structure, the AWS companies concerned, and the patterns you possibly can contemplate when designing your personal multi-agent buyer help resolution.

The problem

As FBG scaled, their present help mannequin required extra human touches per interplay, creating greater operational prices that grew proportionally with their buyer base. The workforce acknowledged a possibility to enhance their prospects’ expertise whereas making ready the help infrastructure for the subsequent part of development. Several components made the issue particularly troublesome.

Every U.S. state has its personal guidelines for cost strategies, deposit limits, withdrawal timelines, and accountable gaming necessities. A buyer in Indiana will get totally different solutions than one in New Jersey. During main sporting occasions, help requests can surge to over 40 inquiries each two minutes, and the system must scale immediately with out degrading response high quality.

The range of queries compounds the issue. Customers ask about the whole lot from transaction historical past and account settings to betting guidelines and self-exclusion choices. No single mannequin or data base covers all of it. On high of that, operators should determine and reply to indicators of downside playing in actual time. This requires a nuanced understanding of conversational context, not simply key phrase matching.

FBG wanted a system that might deal with this complexity autonomously whereas realizing precisely when to escalate to a human agent.

“As we scaled, we knew our support experience needed to evolve with us. We wanted to give our customers faster, more accurate answers while making sure we never compromised on responsible gaming or compliance. The goal was to build something that got better over time, not just bigger.”

— Ian Botts, CTO, Fanatics Betting and Gaming

Solution overview

Rather than counting on a single monolithic chatbot, FBG designed a multi-agent system the place specialised brokers deal with totally different facets of the client interplay. Because the workforce had already constructed deep operational experience on Amazon Elastic Kubernetes Service (Amazon EKS), they may construct on their present container platform to deploy, scale, and iterate on every agent independently.

FBG selected Amazon Bedrock for its model-agnostic entry to a number of basis fashions by way of a single API, which lets the workforce match every job to the most effective mannequin and swap fashions as higher choices emerge. Because Bedrock runs inside their present AWS setting, the system additionally inherits FBG’s established safety and governance controls, and Amazon Bedrock Guardrails gives the accountable AI safeguards their compliance necessities demand.

The structure follows an orchestrator sample. A main orchestration agent receives every buyer message, coordinates with specialised instruments and sub-agents, and returns a unified response. This design permits the workforce so as to add new capabilities, comparable to new instruments, new data domains, and new enterprise items, with out rewriting the core system.

“We designed the system so that each agent has a clear responsibility and can be improved independently. That modularity is what allows us to move fast. When we need to support a new case type or a new business unit, we add a new tool or agent without touching the rest of the system.”

— Luis Fernandez Rocha, Sr. Manager of Software Engineering, Fanatics Betting and Gaming

Figure 1 illustrates the high-level structure.

Figure 1: End-to-end structure of the multi-agent buyer help system on AWS

A buyer message enters by way of the FBG cell app, passes by way of Salesforce Einstein to the Spring AI service on Amazon EKS, then flows by way of Amazon Bedrock Guardrails and a Responsible Gaming classifier earlier than reaching the Supervisor Agent. The Supervisor Agent invokes specialised instruments, together with a Retrieval Augmented Generation (RAG) pipeline, account and transaction Model Context Protocol (MCP) servers, and a transfer-to-agent device to generate a response.

How a request flows by way of the system

A buyer sends a message by way of the FBG cell app, which connects to Salesforce Einstein because the chat interface layer. The request is routed by way of customary REST calls to the Spring AI service operating on Amazon EKS. This service validates the client token and invokes the AI agent.

The request then passes by way of Amazon Bedrock Guardrails to assist detect immediate injection earlier than reaching the AI layer. A Responsible Gaming classification agent, powered by Amazon Nova 2 Lite, evaluates each message in opposition to a compliance-approved classification framework. High-severity classifications set off an instantaneous switch to a human agent with full dialog context.

The Supervisor Agent, operating Anthropic Claude on Amazon Bedrock, determines which instruments to invoke based mostly on the client’s intent. Depending on the question, the supervisor calls a number of specialised instruments, some by way of MCP and others native to the service:

• Retrieval Augmented Generation (RAG) Tool retrieves related data from a vector retailer for FAQ-style questions.
• Account Tools (MCP) queries inside account companies for customer-specific info.
• Transaction Tools (MCP) retrieves current transaction historical past together with deposits, withdrawals, and betting exercise.
• Transfer-to-Agent Tool escalates to a human agent when the client explicitly requests it or when the scenario requires human judgment.

The supervisor synthesizes the device responses and returns a pure language response to the client.

Deep dive: Key architectural parts

In this part, we study the 4 parts that make the system work: the Amazon EKS internet hosting platform, the customized RAG pipeline, the accountable gaming classifier, and the guardrails that assist maintain conversations protected.

Amazon EKS for agent internet hosting and MCP servers

FBG runs their complete AI stack on Amazon EKS internet hosting their MCP server and Spring AI service as Kubernetes companies. The MCP server exposes instruments that make REST calls to exterior companies just like the account service and transaction-history service. Local instruments stay instantly within the Spring AI service alongside the Claude massive language mannequin (LLM). These embrace the RAG device and Transfer-to-Human device.

This method gives a number of benefits for multi-agent methods. The MCP server and Spring AI service scale independently based mostly on demand. When FBG must help extra enterprise domains or options, including a brand new MCP server is simply one other Kubernetes deployment. The workforce also can replace particular person instruments with out redeploying your entire system. New MCP instruments are added to an present MCP server with out requiring new pod deployments.

FBG makes use of Spring AI as its utility framework, which gives native MCP help. The MCP server defines instruments that the supervisor agent can uncover and invoke dynamically. The workforce selected Spring AI as a result of their builders had deep Java experience, which allow them to transfer shortly. For groups working in Python, Strands Agents is an open supply SDK from AWS that gives related agent orchestration and MCP help.

For groups contemplating the same method, Amazon EKS gives the container orchestration wanted to handle a number of agent companies at scale. MCP gives the standardized protocol for device communication between brokers. Teams that choose a managed expertise also can discover Amazon Bedrock AgentCore, a platform to construct, join, and optimize brokers at scale, with any framework or mannequin. AgentCore additionally helps MCP for device integration.

Custom RAG with Amazon Titan embeddings

The most steadily used device within the system is the RAG pipeline. FBG constructed a customized implementation quite than utilizing a managed data base, giving them exact management over the ingestion, chunking, and retrieval course of.

The pipeline works on this approach. Support documentation is collected from upstream sources, together with state-specific cost technique guides, FAQ articles, accountable gaming assets, and account administration guides. Documents are break up utilizing a token-based chunking technique, that means every doc is split into segments of a hard and fast variety of tokens (the items of textual content a mannequin processes) quite than by sentences or paragraphs. This offers the workforce fine-grained management over chunk boundaries. Chunks are then embedded utilizing Amazon Titan V2, producing vector representations saved in MongoDB Atlas.

When a buyer asks a query, the system converts the question right into a vector search-optimized type utilizing an LLM, then performs a similarity search in opposition to the doc retailer. For jurisdiction-specific questions, the system performs each a state-specific search and a normal search, combining the outcomes earlier than passing them to the supervisor agent for response technology.

This customized method is especially worthwhile when your data base has advanced retrieval necessities, like needing to mix state-specific and normal paperwork in a single response. The data base is constantly increasing, with a whole bunch of latest paperwork added month-to-month because the workforce identifies gaps by way of dialog evaluation.

“Building our own RAG pipeline gave us full control over what the model sees and how it retrieves information. Every state has different rules, so we needed the ability to combine state-specific and general documents in a single response. That level of control made all the difference in accuracy.”

— Sharoze Amir, Software Engineer, Fanatics Betting and Gaming

Responsible gaming classification with Amazon Nova

Responsible gaming is a regulatory requirement and a core worth for FBG. The workforce labored with their compliance division to construct a classification system. The system evaluates buyer interactions to substantiate accountable gaming requirements are met and connects prospects with the precise assets when wanted.

The system makes use of Amazon Nova 2 Lite, a light-weight, quick classification mannequin. The workforce selected a smaller mannequin intentionally. The job is well-defined with clear examples and a restricted set of outcomes, so a bigger, dearer mannequin would add latency with out bettering accuracy.

The mannequin receives each the present message and the complete dialog historical past, enabling it to detect escalating patterns quite than counting on single-message key phrase matching. When the system identifies a high-severity concern, it instantly transfers the client to a human agent with full dialog context. Lower-severity flags are recorded for compliance assessment whereas permitting the dialog to proceed.

This is a sample that applies broadly: use the smallest mannequin that meets your accuracy necessities for well-scoped classification duties, and reserve bigger fashions for open-ended reasoning.

“Off-the-shelf support agents treat every conversation the same. Ours can’t — a question about a withdrawal might really be a responsible gaming moment, and recognizing that requires deep integration with our compliance framework. That’s why we built on AWS in-house: no vendor was going to handle those sensitive areas the way our industry demands.”

— Trevor Gurgick, Head of Applied AI, Fanatics Betting and Gaming

Amazon Bedrock Guardrails for safety

FBG makes use of Amazon Bedrock Guardrails to assist shield in opposition to immediate injection and assist maintain conversations inside acceptable boundaries. The workforce tuned their guardrail configuration to assist steadiness safety with the realities of customer support interactions, the place overly restrictive filters can create friction within the buyer expertise.

The key perception: tune your guardrails to your precise use case quite than making use of most restrictions by default. For buyer help, immediate injection safety is essential, however overly aggressive content material filtering creates false positives that frustrate prospects.

A multi-model structure

FBG runs a multi-model structure on Amazon Bedrock, taking a deliberate, bottom-up method to mannequin choice. For classification duties like accountable gaming, they use Amazon Nova 2 Lite. It is quick, cost-effective, and adequate for well-defined classification the place clear examples exist. For supervisor and orchestration duties comparable to dialog administration, they use Anthropic Claude Sonnet on Amazon Bedrock. Claude handles advanced reasoning, device orchestration, and pure dialog. For embeddings within the RAG pipeline, they use Amazon Titan V2, which generates high-quality vector representations. For mannequin availability by AWS Region, confer with Supported fashions by AWS Region in Amazon Bedrock.

The workforce makes use of a round-robin technique throughout mannequin Regions for the supervisor agent, making certain they by no means hit throughput limits throughout peak occasions. Because each mannequin is accessed by way of the identical Amazon Bedrock API, routing totally different workloads to totally different fashions requires no adjustments to the underlying infrastructure.

Results

Within the primary two months of deployment, the multi-agent system delivered measurable enhancements over FBG’s earlier help expertise, based mostly on FBG’s inside metrics. The containment charge improved by roughly 56 %, that means extra buyer points at the moment are resolved with out human agent involvement. Resolution charges improved by roughly 53 %, with prospects getting their issues truly solved quite than deflected. The system has resolved 1000’s of instances autonomously, representing important price financial savings as a result of AI-powered interactions price a fraction of human agent interactions. Customer satisfaction can be trending upward. Conversation high quality has improved so considerably that prospects steadily don’t notice they’re interacting with AI.

During peak sporting occasions, the system handles excessive request volumes whereas sustaining constant efficiency and response high quality. Amazon EKS autoscaling ensures that the MCP servers and Spring AI service add capability routinely as visitors spikes, with out requiring handbook intervention. This implies that whether or not it’s a quiet Tuesday or the Super Bowl, prospects obtain the identical quick, correct responses no matter what number of others are asking questions on the similar time.

Continuous enchancment: Evaluation and iteration

Building the agent is just the start. What makes FBG’s method stand out is the funding they’ve made in constantly bettering the system after deployment. The workforce treats their multi-agent system as a residing product, not a one-time implementation. They assessment dialog logs, monitor decision accuracy, and use actual buyer interactions to refine prompts, regulate device habits, and determine gaps of their data base. This ongoing funding ensures that the system will get smarter over time quite than degrading as buyer wants evolve.

At the core of their analysis technique is an LLM-as-a-Judge system that routinely critiques each accomplished dialog, classifying whether or not the AI efficiently resolved the case or whether or not it fell brief. An operations workforce critiques these evaluations each day, figuring out patterns in the place the agent struggles and submitting enchancment tickets to shut these gaps. This creates a suggestions loop the place the agent will get measurably higher over time.

On the engineering aspect, the workforce displays system well being by way of real-time observability metrics, together with hallucination detection, latency, and value monitoring. When growing new options or testing adjustments, they pull precise buyer conversations from manufacturing and replay them in opposition to the up to date structure to floor edge instances earlier than something reaches prospects.

One self-discipline that has been notably efficient is the workforce’s method to immediate engineering. Rather than leaping to a extra highly effective (and dearer) mannequin when efficiency dips, they first have a look at whether or not the system immediate will be improved or whether or not there are contradictions within the directions. This retains prices low whereas driving steady high quality enhancements, and so they solely improve fashions when the immediate has been absolutely optimized for the duty. Teams trying to systematize this self-discipline can use superior immediate optimization in Amazon Bedrock, which refines prompts in opposition to analysis standards and compares outcomes throughout a number of fashions earlier than committing to a migration.

Getting began: Building your personal multi-agent help system

If you’re trying to construct the same multi-agent buyer help system, right here’s a sensible beginning path.

Define your scope narrowly

FBG launched with simply 4 of their over 20 case varieties. Starting slender allows you to show worth shortly, construct analysis infrastructure, and be taught what works earlier than increasing. Pick the case varieties with the best quantity and clearest decision standards.

Set up your infrastructure

Deploy your agent companies on Amazon EKS for full management or use Amazon Bedrock AgentCore for a managed expertise. Use MCP for communication between your orchestrator and power companies.

Build your RAG pipeline

Start with Amazon Bedrock Knowledge Bases, the absolutely managed RAG functionality, or construct a customized pipeline with Amazon Titan embeddings in the event you want fine-grained management over retrieval logic. Either approach, make investments time in your chunking technique. It has extra affect on response high quality than mannequin choice.

Implement guardrails and compliance from day one

Use Amazon Bedrock Guardrails to assist shield in opposition to immediate injection. If your trade has compliance necessities (gaming, healthcare, finance), construct classification brokers early. They’re simpler to combine from the beginning than to retrofit later.

Start small with mannequin choice

Use the smallest mannequin that meets your accuracy necessities for every job. Reserve bigger fashions to your orchestration agent the place advanced reasoning is required. Amazon Bedrock makes it easy to swap fashions as you iterate.

Invest in analysis early

Build your analysis pipeline alongside your agent, not after. Track containment charges, decision charges, and buyer satisfaction from day one. Use LLM-as-a-Judge patterns to automate dialog assessment at scale.

Conclusion

Fanatics Betting and Gaming’s multi-agent structure demonstrates how combining Amazon EKS, Amazon Bedrock, MCP, and purpose-built AWS AI companies can ship buyer help that scales with your enterprise whereas sustaining high quality and compliance. By utilizing specialised brokers for various duties, together with orchestration, data retrieval, classification, and power execution, the system handles the complexity of state-specific laws, real-time accountable gaming detection, and high-traffic sporting occasions.

The patterns on this put up apply broadly. Define your scope narrowly, use specialised brokers with clear obligations, select the smallest mannequin that meets your accuracy necessities, and spend money on analysis infrastructure from the beginning.

If you’re able to construct the same multi-agent buyer help system, begin by defining your agent boundaries and figuring out the instruments every agent wants. You can deploy your brokers on Amazon EKS for full management over scaling and orchestration or use Amazon Bedrock AgentCore for a managed runtime that handles infrastructure for you. For the AI layer, Amazon Bedrock offers you entry to fashions like Anthropic Claude and Amazon Nova by way of a single API, and Amazon Bedrock Guardrails can assist you add security checks with out customized code. To get began, see the Amazon Bedrock Getting Started information and the Strands Agents documentation for an open-source Python framework that helps MCP-based device orchestration.

About the authors