The hard part of multi-user AI agents isn't the model, it's deciding which slice of a person's private context a shared agent is allowed to see. Instinct's new group chats are a live test of that.
What actually shipped
Instinct is launching group chats where friends use its AI agent together on tasks like planning trips, organizing carpools, and coordinating events. The detail that matters for anyone building in this space: friends without an account can take part. The agent, not a signup form, becomes the interface to the product.
That's a distribution move, and a good one. If a non-user can sit in the thread and still get value from the agent, onboarding happens inside shared context instead of behind a wall. But it also means the agent is now serving people with different context, different trust levels, and different stakes.
Two agents wearing one name
Instinct says personal accounts remain separate, with permission required before personal agents share information or take action. Read that literally and you have two distinct components, not one assistant with a group mode:
- A personal agent holding private context — preferences, history, whatever it knows about you
- A group-scoped agent that only receives what has been explicitly released into the thread
The boundary between them is the product. Everything else — the itinerary, the carpool roster — is downstream of that boundary holding.
Model consent at the thread, not the user
If you model permission as per-user, per-action, a five-person thread turns into a permission-prompt generator. Nobody will approve eleven dialogs to schedule a Saturday. Thread-scoped grants are more workable: this thread gets my availability windows, not my location; this thread can write to the shared plan, not to my calendar.
Revocation has to be visible to the whole thread, not buried in settings. In a group, one person's leaked detail poisons trust for everyone, and a silent revoke reads as either a bug or a betrayal. Make the release of information an event the thread can see.
Why this class of task is the hard case
Planning trips, carpools, and events are coordination problems where the bottleneck is agreement, not generation. An agent that drafts a perfect itinerary the group won't accept has failed. The permission layer is what lets the agent act on someone's behalf without asking them to narrate their whole life to their friends first.
That's the design lesson worth carrying into other multi-user agent products: build the consent gate as a first-class object in the data model, with a clear scope and a visible revocation path. The shared agent's answers are only as good as the context it's allowed to have.
Source: TechCrunch, "Instinct brings its AI agent to group chats, even for friends without an account," Sarah Perez, 2026-10-05.
Top comments (0)