Introducing stateful MCP client capabilities on Amazon Bedrock AgentCore Runtime

Amazon has introduced stateful Model Context Protocol (MCP) client capabilities on its Amazon Bedrock AgentCore Runtime, fundamentally changing how AI agents can interact with users and systems. Previously limited to stateless implementations where each request was independent, developers building agents faced significant constraints when workflows needed to pause for user clarification, request LLM-generated content mid-execution, or provide real-time updates during long-running operations. The new stateful mode solves these limitations by implementing three key MCP client capabilities: elicitation (requesting user input), sampling (requesting LLM-generated content), and progress notification (streaming real-time updates).

This transition from stateless to stateful is enabled by a simple configuration change—setting stateless_http=False—which triggers AgentCore Runtime to provision a dedicated microVM for each user session with CPU, memory, and filesystem isolation. Sessions can persist for up to 8 hours with a 15-minute idle timeout, maintaining continuity through session IDs. The three new capabilities transform one-way tool execution into true bidirectional conversations: elicitation allows agents to ask targeted questions at precise moments in workflows, sampling enables dynamic content generation from the client's LLM, and progress notifications keep users informed during lengthy operations. This completes Amazon's bidirectional MCP protocol implementation, moving beyond the previous stateless server hosting capabilities.