Change the model. Keep ownership of the work.
The agent and its model have different roles
A model supplies reasoning for an agent’s response. The continuing agent has its own identity, objective, state, work and authority. Keeping those responsibilities separate allows work to remain attached to the right participant rather than to a particular provider’s chat session.
Changing models should not claim that the new model performed earlier actions or invalidate the authorship of retained outputs. The work still needs its original context and observed results. A provider’s ability to generate an answer also does not grant access to any system the answer describes.
Bring organization-approved model access
Installed Cuttlefish supports model connections supplied and authorized by the customer or organization. Model availability depends on those connections, provider access and applicable policy. Cuttlefish should not be read as supplying bundled access to every advertised model.
An organization can decide which connections its agents may use. Those choices can reflect confidentiality, operational requirements and provider arrangements. A personally available connection is not automatically approved for organizational work, and an agent must not infer approval from a model’s display name.
Assign a model to the right responsibility
Companions can have their own permitted model assignments. This lets distinct responsibilities use suitable authorized connections while remaining part of the same objective. The assignment is part of the agent’s operating context, not a claim that all models are interchangeable in quality or supported behavior.
Each active model response uses its assigned connection. An authorized change applies to subsequent continuation within the product’s supported behavior. A pinned model must not be silently replaced because another provider is easier to reach or a request has failed.
Continuity still has practical limits
An agent can retain its work while changing an authorized model connection, but a new model response is not a copy of another model’s private internal state. The continuing response depends on the relevant retained information available to it and the capabilities of the selected provider.
Different models can have different context limits, tool behavior and operational constraints. Provider independence means preserving product ownership and permitted choice; it does not promise identical outputs, infinite context or automatic compatibility with every model endpoint.
Unavailable inference is an explicit state
A connection can be unavailable, throttled or refused by policy. The system should identify that outcome rather than invent a successful response. Any retry or permitted reassignment must keep the current work identity and policy in view.
Usage and cost also have a specific owner. Installed customers’ provider usage follows their actual connections and arrangements. The public website’s configured inference and visitor limits are a separate operating boundary; they are not a statement about an organization’s native-product model budget.
The website uses its configured model
This website does not offer a public model selector. Its Home and Companions use the website’s configured capabilities and bounded example information. Asking the website to connect a private provider does not create an installed-product connection.
For evaluation, check the intended provider, permitted model assignments, information handling and failure behavior in the actual installed environment. Enterprise and the controlled Privacy and Subprocessor documents provide the appropriate context for organizational review.