Build Interactive Multi-Turn Agents Using Stateful MCP Sessions
AWS · Feature Update · · notable
Briefing for: Engineering
What happened
Amazon Bedrock AgentCore Runtime now supports stateful Model Context Protocol (MCP) features, including elicitation, sampling, and progress notifications. Sessions are isolated within dedicated microVMs and use the `Mcp-Session-Id` header to maintain context across multi-turn interactions.
Why it matters
This enables complex agentic workflows where servers can pause execution to ask for user input or LLM-generated content without losing session state. The introduction of progress notifications also allows for better handling of high-latency operations like data processing or external API calls.
What this enables
- If you build complex RAG workflows, use elicitation to have the agent clarify ambiguous user queries mid-execution.
- If your agent performs long-running tasks, implement progress notifications to keep the client UI responsive and avoid timeouts.
- If you need AI-powered sub-tasks during tool execution, use sampling to request generation from the client-side LLM.
Get personalized AI briefings for your role at Changecast →