Multi-Agent Systems for Financial Services on Amazon EKS and AgentCore

This article demonstrates how to build production-grade multi-agent systems for financial services using Amazon EKS, Amazon Bedrock, and Amazon Bedrock AgentCore with enterprise-grade authentication, authorization, tracing, and secure code execution.

• Multi-agent architecture separates agents into narrow domains with an orchestrator coordinating specialists for tasks like portfolio analysis, risk assessment, and market data gathering
• Amazon EKS Auto Mode manages compute, networking, and security while agents run as Kubernetes Deployments with RBAC and Pod Identity
• Amazon Bedrock provides foundation model inference (Claude Sonnet/Haiku) with data remaining within AWS account boundaries
• Amazon Bedrock AgentCore provides Memory for cross-session context, Code Interpreter for safe Python execution, and Browser for live market data
• Agent Gateway enforces authentication via EKS OIDC JWTs and authorization via CEL policies for both agent-to-agent and tool calls
• Model Context Protocol (MCP) standardizes tool definitions; LiteLLM provides central LLM routing with rate limiting, caching, and cost tracking
• GitOps with ArgoCD and Crossplane enables declarative agent lifecycle management and reproducible platform deployments
• Reference implementation available on AWS Samples GitHub with Terraform, agent code, and bootstrap scripts for 30-minute setup

The architecture provides a blueprint for financial services organizations to build scalable, auditable multi-agent systems with production-grade security, observability, and cost controls.