[Agentproto] Re: Topic: Existing practice, use cases, and deployment models

[Nicholas Templeman <[e-mail]>]

Hi Orie, Leslie and all,





I would use draft-agentic-ai-usecases-requirements-03 as the starting point, with a short traceability mapping to Rosenberg-00. Its dialog/task-context distinction and numbered requirements fit the charter. Rosenberg's section 3.3.3 remains useful for pending downstream operations after reconnection, and section 3.3.5 for delegated authorization, without bringing application context or user-interface requirements into the dialog protocol.





Tiru's proposed -04 text addresses the current-dialog checks, refusal when checks cannot complete, acceptance-time authorization, and the distinction between resuming a dialog and learning an operation's outcome. I support retaining those clarifications.





The remaining precision I support is Bradley's sentence: "The authorization context may be carried in the dialog or by reference." At the lifecycle-transition decision, the participant should also distinguish a reference it could not resolve in time from a resolved value outside the applicable freshness bound, and distinguish both from an evaluated denial. These describe the inputs available for evaluation; they are not additional authorization decisions.





That is a targeted clarification to the proposed starting point. It need not prescribe an audit format, transport task context, or replace authorization mechanisms developed with the relevant working groups.





Best,


Nicholas