A request is evaluated in the context of the user, tenant, workspace, connected capability, resource, and policy. A tool merely being discoverable does not make every operation permitted.
Governance for action
An action should be authorized, observable, and accountable.
Getting an answer is one kind of work. Changing a customer record, notifying a team, or triggering an external workflow is another. Spine Runtime helps preserve that distinction.
Intent is not permission
Approval where it matters
For actions requiring explicit human authorization, the Workbench presents a proposal for review before a governed command is issued. Approval is tied to the intended action and scope.
Execution is not the end of the story
Observe the provider response, correlate it with the original work, verify supported readback, and distinguish confirmed, failed, and unresolved outcomes. The resulting continuity is useful only when it reflects what actually happened.
Built for bounded delegation
An AI agent cannot legitimately gain more authority simply because it calls another agent or tool. Policies and connection bindings define the ceiling of permitted actions.
Continue with context