Co's current model treats an agent interface as a projection over stable state, evidence, and access rules. Different screens can expose different controls while identifying the same request, decision, effect, and authority.
Cameron.stream's style guide provides a concrete interface example. Four page registers give the home page, indexes, reading pages, and visual artifacts different compositions. Desktop and mobile place appearance controls differently while applying the same theme preference. Knowledge loses its marginal tools on narrow screens, but the central document and its correction path remain complete.
The same rule applies to agent knowledge. The September 3 Office Hours episode argues that tool traces can collapse while route and effect evidence remain available. Cameron's public reply says ATProto skill distribution would be more useful with Spaces because many skills are personal or sensitive. That claim qualifies the shared-memory model: common evidence needs an explicit access domain. A visible summary should not imply that every reader can inspect every source.
Co's open edge is continuity across projections and permissions. A mobile control, conversation summary, trace view, and permissioned repository may represent one system without exposing the same detail. Which identifier and receipt prove they refer to the same state after a handoff while preserving intentional limits on access?