LangGraph 에이전트를 Bedrock AgentCore로 이관한 PoC 후기

The recent migration of agentic workloads to Amazon Bedrock AgentCore was explored in this post. While running the LangGraph-based customer support agent demo on a laptop, it became apparent that the session isolation, checkpoint persistence, and tool authentication were all part of our own code once real users started engaging. Amazon Bedrock AgentCore is a platform that removes operational burdens from agent inference, handling runtime, tools, and memory separately.

The runtime is replaced by microVMs for each session, the Gateway acts as an MCP (Model Context Protocol) tool publisher, and Memory maintains conversation state even if the process crashes. The source material divides this into Stage 1 (Stage boundaries), Stage 2 (loop redesign using model-driven planning), and Stage 3 (harness).

This article presents the migration design by examining the original source and public sample repository, aws-samples/sample-migrate-agents-to-amazon-bedrock-agentcore, with a real deployment in my AWS account to verify the findings.

The post outlines the core claims, which include 45/22/85 lines changed in the agent, adapter code, and stage 0 imports, respectively. Gateway tool names start with '___' prefixes, and cross-process memory continuity is maintained. The deployment confirmed OTEL settings are unnecessary, generating CloudWatch logs automatically. However, there are caveats to consider, such as potential issues with Mac ARM or x86 vendorizing, and the unreliability of event counts for asserting continuity.

The Stage 2 trade-off is the loss of deterministic routing, as the model-driven planning removes a hard-coded branch. Ultimately, Stage 1 is recommended for most teams due to the lack of hard-coded bottlenecks in the LangGraph graph and the absence of new intents that would require prompt handling.