Summary
- • AWS adds stateful MCP support to Bedrock AgentCore Runtime, enabling bidirectional multi-turn agent workflows
- • Three new client capabilities: elicitation (mid-run user input), sampling (LLM content requests), and progress notification
- • Each session gets a dedicated microVM with isolated CPU/memory/filesystem, persisting up to 8 hours
- • Completes the full bidirectional MCP protocol implementation on AWS, replacing stateless-only mode
Details
Stateful MCP client capabilities now available on Amazon Bedrock AgentCore Runtime
Extends prior stateless MCP server hosting to support the full bidirectional MCP specification, enabling server-to-client requests within a persistent session. Activated via stateless_http=False.
Elicitation capability lets MCP servers pause mid-execution and request user input
Previously impossible in stateless mode. Enables agents to ask clarifying questions during tool calls rather than failing or making assumptions when ambiguous inputs arise.
Sampling capability allows servers to request LLM-generated content from the client
Servers can offload generation tasks back to the client-side model rather than handling all inference internally, aligning with the MCP spec's client capability model.
Progress notification streams real-time updates from server to client during long-running operations
Fills a major gap for agentic workflows where operators or users need visibility into task status without polling or waiting for final output.
Stateful mode provisions a dedicated microVM per user session with full CPU, memory, and filesystem isolation
Sessions persist up to 8 hours, with idle timeout configurable through idleRuntimeSessionTimeout setting (default 15 minutes of inactivity).
Session continuity managed via a Mcp-Session-Id header returned during the initialize handshake
The client includes this header in every subsequent request, routing traffic back to the same microVM instance for the duration of the session.
Stateless mode remains available and appropriate for simple input-output tool servers
The two modes serve different use cases: stateless for straightforward request-response tools, stateful for interactive or long-running agentic workflows requiring mid-execution communication.
Technical capabilities, implementation details, and deployment implications of stateful MCP on Amazon Bedrock AgentCore Runtime
What This Means
AI practitioners building production agents on AWS can now implement workflows that pause for user input, stream progress, or delegate generation back to the client — all patterns that stateless infrastructure made impossible. The dedicated-microVM-per-session model provides the isolation and persistence needed for complex multi-turn agent tasks without managing that infrastructure manually. Teams already using AgentCore Runtime for tool hosting can upgrade to stateful mode with a single configuration change (stateless_http=False), though they should account for the 8-hour session limit and 15-minute idle timeout when designing longer-running workflows.