Migrate agentic workloads to Amazon Bedrock AgentCore
Amazon Web Services Sruthi Vedula ● Covered by 3 sources
AWS shows how to move an agent from LangGraph to Bedrock AgentCore. The big win is cutting ops work like patching, tool auth, and session state.
Based on reporting by Amazon Web Services, Sruthi Vedula — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a practical migration path for agents that already work in a notebook or a local app, but fall apart once real users show up. The problem, in its telling, isn’t the reasoning loop. It’s everything around it: keeping sessions apart, preserving state over time, wiring tool auth, and patching the machine underneath.
The walkthrough starts with a LangGraph customer support agent that already calls Amazon Bedrock for inference. That matters, because it means the move is not about swapping models. It’s about moving the plumbing. In stage 1, the graph stays intact while Amazon Bedrock AgentCore Runtime takes over execution, Gateway fronts two tools, and Memory takes the conversation state. The agent still classifies a message, escalates angry customers, and uses three tools; only the operating burden shifts.
AWS breaks the migration into stages on purpose. Stage 0 is the baseline: a compiled StateGraph with a hand-written router, three tool functions, and a MemorySaver checkpointer keyed by thread_id. Stage 1 wraps that same graph in BedrockAgentCoreApp, pulls the session ID from the runtime context, and keeps the rest imported from the original code so the behavior doesn’t drift. The post says 45 lines change inside the agent, 22 lines are supporting code, and 85 lines are imported untouched.
Stage 1 also moves two tools behind AgentCore Gateway as MCP tools published from Lambda. lookup_order and process_return are registered there; search_faq stays local. Gateway takes over tool auth and invokes the Lambda under its own execution role, while the agent discovers the tools over SigV4-signed MCP calls. Memory shifts from a process-local checkpointer to AgentCoreMemorySessionManager, which is the piece that keeps state across turns, processes and days.
The source also keeps the limits in view. Runtime does not replace IAM, VPC setup, WAF rules, or secrets rotation. Those stay with you. Stage 2, which the post points to after this first move, rebuilds the loop around model-driven planning on Strands Agents. Stage 3 is mentioned as documented, not built. The point of the whole exercise is plain: move the work that has nothing to do with the agent’s actual reasoning.
My take — AI-written commentary, not fact-checked reporting
This is the rare cloud migration story that sounds boring in exactly the right way. If an agent needs a private server, a hand-rolled checkpointer, and tool auth glued into app code, it’s already doing ops cosplay. The real pattern here is that “agent platform” only becomes meaningful when it absorbs the unglamorous bits and leaves the model to do model things.
Read more about this at: Amazon Web Services