Running Claude Agent SDK on Amazon Bedrock AgentCore

Thu, 02 Apr 2026 12:26:46 +0000

The application I’m building has a chat interface where users guide an agent through multi-step tasks. Four requirements make it non-trivial:

Streaming — responses arrive token by token, not as a single payload after the agent finishes thinking
Human-in-the-loop — the agent pauses mid-execution and asks the user to approve or deny a tool call before proceeding
Persistent sessions — conversation state survives container restarts and can be resumed exactly where it left off
Isolation — each user gets a completely separate execution context; one user’s session cannot bleed into another’s

For the agent, I’m using the Claude Agent SDK — which wraps Claude Code, the same agent behind the Claude Code CLI. Claude Code ships with a built-in tool harness: file editing, bash execution, web search, code analysis. With LangGraph you’d define every tool yourself. Beyond the harness, coding agents are general-purpose agents in disguise — anything a computer can do can be framed as a coding task, which gives Claude Code a natural interface to nearly any tool. As of early 2026, it’s the strongest agent operating in that space.

For the runtime, I’m using Amazon Bedrock AgentCore. The choice of framework and runtime shapes how hard each of the four requirements above is to build — and understanding that is what this post is about.

This post is about running the Claude Agent SDK on Amazon Bedrock AgentCore, and specifically about why the SDK’s execution model is different from LangGraph or Strands in ways that matter for this problem.

AI Engineering on AI in Production

Running Claude Agent SDK on Amazon Bedrock AgentCore