Repose Agent demonstrates how AWS’s new Bedrock AgentCore framework enables the next generation of intelligent, secure, and scalable conversational AI systems. Built by ZirconTech, the agent helps users find funeral services tailored to their needs — a use case demanding contextual memory, real-time tool invocation, and strict governance.
The Journey to a Production-Ready Agentic System
AgentCore Foundation
The conversational agent runs on AWS Bedrock AgentCore, with Claude 3.5 Haiku as the underlying LLM. The runtime is containerized, deployed through AWS CodeBuild pipelines, and stored in Amazon ECR for managed versioning. The design follows a serverless pattern, using Bedrock’s built-in scaling to handle unpredictable query loads without manual provisioning.
Stateful Conversations with Memory
The agent integrates AgentCore Memory using a session-based model that retains dialogue context for seven days. Each session stores summaries under a structured namespace (repose/conversations/session/{sessionId}), retrieved dynamically at the start of every invocation. This approach gives the agent continuity across conversations while keeping operational costs low.
Dynamic Tool Provisioning via MCP
A major innovation in this system is the Agent Gateway, built on the Model Context Protocol (MCP). The gateway automatically generates tools from an OpenAPI specification stored in S3, exposing backend APIs as callable functions for the agent. Updating the OpenAPI schema instantly extends the agent’s capabilities — a zero-code update pattern that allows new features to go live in minutes.
Secure Authentication and Configuration
Service-to-service authentication uses Amazon Cognito with an OAuth 2.0 machine-to-machine (M2M) flow. Tokens are refreshed dynamically by a factory pattern during long sessions. All environment variables, credentials, and URLs are stored in AWS Systems Manager Parameter Store, ensuring configuration parity across Dev, QA, and Prod without code changes.
Fine-Grained Governance
Two IAM roles enforce least-privilege access — one for the Agent Runtime, another for the Gateway — each restricted to Bedrock, ECR, CloudWatch, X-Ray, and Parameter Store operations. Logging and tracing are centralized through CloudWatch and AWS X-Ray, giving engineers full observability across every interaction.
Architecture Overview
Frontend: Next.js interface → Route handler invoking Bedrock AgentCore runtime via AWS SDK
Backend: Agent Gateway using MCP protocol to connect to REST APIs derived from OpenAPI schema in S3
Memory: AgentCore Memory service maintaining contextual state per session
Security: Cognito JWT auth, IAM roles, SSM parameters, organization-wide controls
Results and Impact
- Zero-code extensibility: new APIs exposed as tools within minutes by updating the OpenAPI schema in S3
- Persistent, contextual conversations across user sessions with summary memory strategy
- Full observability: CloudWatch Logs, Metrics, and X-Ray traces track agent latency, tool calls, and gateway interactions
- Secure operations: multi-layer authentication (Cognito + API Key), least-privilege IAM, centralized configuration
- Scalable runtime: containerized Bedrock AgentCore deploys automatically with CodeBuild/ECR pipelines
Lessons Learned
- Memory design defines experience — summary-based memory balances context depth and cost.
- MCP protocol standardizes extensibility — external APIs can be added without retraining or redeployment.
- Centralized configuration wins — Parameter Store keeps deployments environment-agnostic.
- Observability must be built in — X-Ray and custom metrics make agent behavior explainable.
- Automation ensures repeatability — CodeBuild pipelines and IaC keep environments consistent and auditable.
Final Thoughts
Repose Agent showcases a stateful, secure, and fully managed conversational AI system powered by AWS Bedrock AgentCore. By merging agentic orchestration, MCP tooling, and strong governance, ZirconTech delivered an architecture ready for real-world workloads — a blueprint for teams adopting AWS Bedrock Agents to power intelligent, compliant, and scalable customer interactions.