Independent agents, accountable collaboration.
Identity outlasts one model response
A Companion is an independently executing agent assigned to work. Its identity distinguishes its objective, state, conversation, tools, model connection and authority from those of Home and other Companions. Naming several chat threads does not provide these operating relationships by itself.
The assigned model contributes reasoning to a turn. The Companion owns the continuing work. Changing an authorized model connection for subsequent work should not reattribute earlier findings or turn another participant’s outputs into the new model’s accomplishments. Provider availability and policy still constrain what can continue.
Bounded assignments make parallel work useful
Home can separate an objective into responsibilities and supervise their progress. A useful assignment identifies the intended result and the context the agent can use. Different participants can investigate or create in parallel when their dependencies allow it, while decisions requiring a shared result remain coordinated.
The person can also direct a Companion. Follow-up belongs to that participant and its continuing assignment. Starting new work does not silently authorize additional systems, copy another guest’s state or discard an unresolved earlier action. Work that is blocked or needs a decision should be visible as such.
Agents can communicate directly
Companions can exchange authorized findings directly. The person does not have to copy each message between agents. Addressed communication and bounded shared context let a participant ask for relevant work or use a result while preserving who supplied it and what it concerns.
A shared mailbox is a communication surface for permitted participants, not a source of universal authority. Receiving a finding does not confer the sender’s credentials or unrestricted access to the original source. A result can become another participant’s starting point only within the receiving agent’s own scope.
Composite Companions coordinate a shared responsibility
A Composite Companion brings cooperating agents together for work that benefits from several distinct responsibilities. The coordinating responsibility and individual member work remain meaningful: membership does not collapse all state, model assignments or permissions into a single unrestricted agent.
Judge the collective result against the requested objective and the member outcomes it depends on. A member may finish while another fails, waits for permission or needs more information. Partial progress should not be presented as complete collective success. The same limits apply when cooperating work is represented in a Working Group.
Keep authorship and permission with the result
A result can be presented beside Home’s response, reused in a surface or included in shared work. Its author remains the originating agent. Presentation is not a transfer of ownership, and a copied description is not a substitute for the actual retained artifact.
Revocation and changed scope can remove access to information or prevent later work. They do not make an already completed external action disappear. When a participant becomes unavailable, the remaining agents must distinguish a retained permitted result from material that is no longer accessible.
What a visitor can try here
Website Companions operate on the public environment’s available example information and permitted context. Their activity can produce real on-site responses and work products, but Juniper Coffee is fictional and does not represent access to a visitor’s business.
The website cannot open a visitor’s private applications or invite external people into an organization. Installed-product collaboration uses the organization’s actual memberships, connections and policy. Use the Working Groups and Enterprise references to evaluate that boundary before treating a public demonstration as a deployment commitment.