Theo Browne explains that the original Model Context Protocol required a dedicated stateful connection between every client and server. Even unused tools could create persistent processes and connection overhead, especially when one coding-agent session launched several subagents.
The new stateless core replaces that relationship with request-response calls. MCP servers can run on serverless or edge infrastructure, scale without sticky routing and expose self-describing tools that are easier to audit than unrestricted shell access. Theo Browne sees this as a major improvement for sensitive agent workflows.
Compatibility is the remaining concern. Older clients and servers may claim MCP support while implementing the stateful protocol, so the new design is not automatically interoperable with every existing integration. Theo Browne argues that agent-tool vendors should update quickly and make their supported protocol version clear.
Watch on YouTube
