Amazon Web Services has launched stateful MCP client capabilities on Amazon Bedrock AgentCore Runtime, allowing developers to build and deploy AI agents that maintain persistent session context, interact with users during execution, and stream progress updates across long-running tasks.

The Model Context Protocol, an open standard originally introduced by Anthropic to standardise how AI models connect to external tools and data sources, has until now been largely implemented in stateless configurations. Stateless MCP servers handle each request in isolation, which limits the complexity of agentic workflows that require memory, mid-task user interaction, or dynamic content generation across multiple steps.

Why Stateful MCP Changes the Agent Development Picture

The new capabilities address three specific gaps in standard MCP implementations. First, developers can now build MCP servers that pause execution to request user input — enabling human-in-the-loop workflows without abandoning the broader task context. Second, agents can invoke LLM sampling dynamically during execution, allowing for on-the-fly content generation within a running workflow rather than requiring all model calls to be pre-planned. Third, long-running tasks can stream granular progress updates back to the client in real time.

Stateful MCP servers that pause for user input, invoke LLM sampling mid-task, and stream progress updates represent a shift from tool-calling to agentic behaviour.

Taken together, these capabilities move MCP implementations toward agentic behaviour — where an AI system can reason, act, wait, and adapt — rather than functioning as a sophisticated but static function-calling layer.

What Developers Can Actually Build

According to AWS, the release includes working code examples for each capability, alongside a deployment path directly to Bedrock AgentCore Runtime. Developers can deploy stateful MCP servers without managing the underlying infrastructure for session persistence or streaming, as the Runtime handles those concerns at the platform level.

Practical use cases include multi-step data processing pipelines that require human approval at checkpoints, document generation workflows that call the LLM iteratively based on intermediate outputs, and monitoring or analysis tasks where real-time progress visibility matters to end users or downstream systems.

The integration path uses standard MCP tooling, meaning developers already familiar with the protocol can adopt stateful features incrementally rather than rebuilding from scratch. AWS has not published separate pricing for stateful MCP sessions specifically; costs follow Bedrock AgentCore Runtime pricing, which is based on compute and invocation metrics.

Bedrock AgentCore Runtime's Role in AWS's Agentic Stack

Bedrock AgentCore is AWS's managed execution environment for AI agents, designed to handle the operational complexity — security, scaling, observability — that comes with deploying agents in production. The Runtime component specifically targets the execution layer, where agents call tools, manage context, and interact with external services.

Adding stateful MCP support positions AgentCore Runtime as a more complete platform for production agentic workloads, competing with similar infrastructure layers from Microsoft Azure (via Azure AI Agent Service) and Google Cloud (via Vertex AI Agent Builder). The MCP standard itself is gaining broad adoption, with support from OpenAI, Anthropic, and a growing ecosystem of third-party tool providers.

The stateful capabilities also reduce a common workaround in current agentic architectures, where developers add external session stores — databases, caches — to otherwise stateless MCP servers. By handling state natively in the Runtime, AWS reduces both engineering overhead and potential failure points in production deployments.

What Happens Next for MCP on AWS

AWS has not announced a public roadmap for further MCP enhancements on Bedrock, but the pattern of this release — closing gaps between the MCP specification's theoretical capabilities and real-world deployment constraints — suggests continued investment in the protocol as a primary integration layer for its agentic infrastructure.

For developers evaluating where to run production agents, the combination of managed state, streaming, and LLM sampling invocation within a single Runtime reduces the number of external dependencies required to ship a reliable agentic system. The availability of code examples at launch lowers the barrier to evaluation for teams already using Bedrock.

What This Means

Developers building multi-step AI agents on AWS can now implement persistent session state, mid-task user interaction, and real-time streaming natively within Bedrock AgentCore Runtime, removing the need for external session management infrastructure and making complex agentic workflows significantly more practical to deploy in production.