Part 3: Building the AI Agent with Strands Agents SDK, Prompt Caching, and AgentCore Memory

With the CDK infrastructure in place (Part 2), we need an actual agent to run inside it. The agent is a Python application that: Exposes an HTTP endpoint AgentCore can call Uses the Strands Agents SDK to run a Bedrock-backed reasoning loop Integrates with AgentCore Memory for persistent context Uses Bedrock Guardrails on every invocation The full source is in apps/customer-service-agent/ in the demo repo. Why Strands over LangChain or LlamaIndex? When I started this project, LangChain was the default answer for "I need to build an agent." I used it, ran into friction, and switched to Strands. Here's why: Strands is AWS-native. It's built to integrate directly with Bedrock services — prompt caching, guardrail configs, tool definitions. With LangChain, you write adapter code to bridge from LangChain abstractions down to raw Bedrock APIs. With Strands, you're calling the Bedrock API directly through a thin, intentional abstraction. Tool definitions are simpler. In LangChain, you define to