A capable component should not automatically become an authoritative component.
VAJRA separates model reasoning from policy and execution authority. GeminiAgentBridge applies the same idea at a smaller boundary: the reasoning backend may propose a tool call, but the downstream agent remains responsible for execution.
This distinction is useful beyond AI. Storage sidecars, information pipelines and automation systems also become easier to reason about when derived capability cannot silently redefine the source of truth.
The practical test is simple: if this component is compromised, wrong, or replaced, what prevents it from changing what the system considers authoritative?
A good architecture has a concrete answer.