AI control moves from a policy appeal to the tool call
As 1,134 industry employees seek shared tools for pacing development, MCP is moving responsibility for state and control out of the protocol and into server implementations.
Artificial Intelligence··Morning
Employees seek a shared way to set the pace
A statement signed by 1,134 employees of OpenAI, Anthropic, Google and Meta asks the US government to support an international effort to develop technical and governance tools for deliberately pacing automated AI development. The signatories describe capability growth outrunning the ability to understand or control resulting systems as a risk. Backed by Guidelight AI Standards and Encode AI, the appeal seeks shared tools for managing the pace rather than announcing an immediate halt.[1]
MCP leaves state to the application
MCP's 28 July revision removes protocol-level sessions and the Mcp-Session-Id header from the Streamable HTTP transport, along with the initialize and notifications/initialized handshake. Each request now carries its version and client capabilities in _meta, while servers must advertise supported versions and capabilities through server/discover. When state is needed across calls, the protocol no longer holds it; the server mints its own handle and passes it as an ordinary tool argument.[2]
The same control question at two layers
The developments are not implementations of one another: one concerns coordination among governments and companies, the other responsibility among software components. They nevertheless expose the same question: pace and state do not manage themselves, so an actor must explicitly own them. The statement looks for that ownership in international technical and governance work. MCP gives state to the server author and opens a minimum twelve-month deprecation window for Roots, Sampling and Logging, leaving implementers with a concrete migration burden.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
