Scenario 2: Me, my agents and guests

Who: the owner, their agents and 1 to 5 guests (partner, editor, collaborator, advanced student). A guest can bring their own agent, as in the video.

A typical day - The guest joins the #project-x channel and asks the owner's agent about progress. - The agent answers with what it knows about the project, without showing the rest of the owner's machine. - The guest connects their Claude, and the two agents split the task (one writes, the other reviews). - Anything that goes outside (publish, send, delete) waits for the owner to approve in the thread.

What changes compared with scenario 1 - Who can do what now exists: the guest talks, but cannot make the owner's agent execute anything. - The guest sees only the channels they were invited to. - Cost: each agent runs on its own owner's subscription, as in the video.

Possibilities - 2a: the guest only talks, and the owner's agent answers in read-only mode. - 2b: the guest can request tasks, and the owner approves each one. - 2c: guest with their own agent, in a shared channel with a cap on turns between agents.

Risks: prompt injection coming from the guest's message (question 3) and privacy of the owner's memory (question 6).

Makes sense when: there is a partnership or a two-person project and the exchange today happens over WhatsApp, with screenshots.