AWS Bedrock AgentCore: stateful MCP client enables interactive AI workflows
Why it matters
Amazon has extended Bedrock AgentCore Runtime with three new MCP capabilities — elicitation (requesting structured input from the user), sampling (requesting LLM completions from the client), and progress notifications. Stateful sessions can now last up to 8 hours in isolated microVMs and enable two-way communication between agent and client.
On April 9, Amazon extended the capabilities of its Bedrock AgentCore Runtime by introducing three new client-side features for MCP (Model Context Protocol). The shift from a stateless to a stateful architecture enables an entirely new class of interactive workflows between AI agents and end users.
What stateful mode enables
In the new stateful mode, AgentCore allocates a dedicated microVM per session that can last up to 8 hours with a 15-minute idle timeout. Previously, stateless servers could not maintain context between requests nor request additional information from the user in the middle of execution. Now true two-way conversation is possible — an agent can pause, ask the user something, receive an answer and continue where it left off.
Three new capabilities
Elicitation lets servers pause execution and request structured input from the user via the client. It supports both form-mode (fields for the user to fill out) and URL-mode (opening an external page). The user can respond with three actions: accept, decline or cancel (aborting the entire operation).
Sampling allows the server to request an LLM completion from the client — the server sends a prompt to the client, the client forwards it to the connected LLM, and returns the result. Notably: the server can express preferences about which model to use (e.g. “I need a strong coding model”) without having direct access to those models itself. This cleanly separates responsibilities — authorization and billing for models remain with the client.
Progress notifications use fire-and-forget messages during long-running operations. The client can render a progress bar, status indicator or animated feedback while the server processes a multi-step action. This solves a UX problem that previously required various workarounds.
In practice, this brings MCP out of the “single-call tool” domain into the domain of “infrastructure for interactive agentic applications”. Combined with the previously announced AWS Agent Registry, this outlines a clear AWS plan for enterprise AI agents.