AWS has detailed how the latest Model Context Protocol (MCP) specification removes protocol-level sessions for remote server deployments, allowing requests to hit any available server instance without sticky sessions or shared session stores. This architectural update simplifies horizontal scaling by shifting state management responsibilities directly to surrounding infrastructure.
Dismantling the Protocol Handshake and Sticky Routing
The updated MCP specification eliminates the legacy initialize and initialized handshake alongside the Mcp-Session-Id header. By stripping out these protocol requirements, incoming requests can now route independently to any server instance sitting behind a conventional load balancer.
AWS Architecture Blog authors Anand Komandooru, Steven DeVries, and Haleh Najafzadeh point out that this change replaces session-affine routing with standard request routing. It also removes the dedicated session storage mechanisms previously required solely to maintain MCP protocol state. The underlying architecture now draws a sharp line between protocol state and application state.
Writing about the specification on LinkedIn, Michael Madsen summarized the architectural paradigm shift with stark clarity: “The protocol is stateless. Your application doesn’t have to be.”
Unlocking Serverless Deployments with AWS Lambda
The removal of persistent session connections changes the calculus for hosting agentic tool servers. AWS Lambda has emerged as a viable deployment option because the updated request-response model no longer demands that infrastructure hold open streaming sockets.
To support multi-step interactions without held-open streams, the specification introduces Multi-Round Task Requests (MRTR). MRTR replaces server-initiated requests by utilizing input_required responses followed by subsequent client requests.
New routing controls have also been embedded into the protocol headers. Specifically, Mcp-Method and Mcp-Name headers enable advanced gateway routing and throttling, while W3C Trace Context support brings native distributed tracing to agentic workflows. Furthermore, developers can leverage ttlMs and cacheScope to handle granular caching controls across distributed calls.
Migration Paths and Idempotency Challenges
Stream resumability has been removed from the specification. Consequently, clients must now handle retries for interrupted operations, placing a heavy premium on tool call idempotency when operations produce side effects.
Existing enterprise infrastructure requires a managed transition path. For instance, Apify’s MCP server project is actively implementing stateless support alongside its existing sessionful server, maintaining routing and conformance tests for both protocol versions.
Migration remains essential for environments supporting earlier MCP clients. AWS recommends tracking protocol versions explicitly at the API gateway layer and retaining legacy session infrastructure until all legacy traffic has been fully phased out. The MCP project has established a formal feature lifecycle policy to provide a defined migration window for any deprecated capabilities.
AWS maps these modernizations directly to its Well-Architected Agentic AI Lens, offering formal guidance on monitoring, tracing, security, and tool integration as teams transition away from sticky sessions toward truly elastic, stateless agent deployments.
