Agents: expose the application, do not reimplement it

A Foldkit application already has the two things an agent needs:

Model         -> information the agent may observe
Message union -> actions the agent may cause

foldkit-agent turns deliberate subsets of those into one protocol-neutral agent contract. WebMCP, MCP, A2A, and Agent Native then serve that same contract without adding application behaviour of their own.

The important architectural rule is simple:

An agent capability is an application Message, not a second implementation of the feature.

The UI and the agent therefore meet at the same update. If a person and an agent both rename a todo, there is still one rename transition to understand, test, authorize, replay, and debug.