The bench: one runtime, three windows
The same question, asked in three different places, reaches the same governed runtime.
The three windows are the Ardinois chat workspace, a Copilot Studio agent and a Foundry agent. Each is a way in. None of them owns the loop that does the work. That loop, with its tools, its knowledge and its records, runs once, in the governed runtime behind all three.
Surface is not runtime
The window a person uses should not decide where the loop, the tools, the knowledge or the state run. The bench makes that claim testable. Ask the same question in each window and the three executions share one correlation identifier family and one runtime version, make the same tool calls and each leave an evidence record.
The bench also shows what is lost when a surface is allowed to drag a runtime along with it: a different loop, different logs and a different set of controls for every product.
The layer lens
Every layer of the architecture is a working component, and each has a lens in the workspace that shows what that layer did for the current request. Where it is safe, a layer can be switched off or swapped during a session so that the audience sees the consequence.
The platform governs itself
Agents and skills built in the workspace pass through the same classification, release levels, identity and trace controls that the architecture prescribes for any capability. Deploying a skill or promoting an agent needs an approval. That is the demonstration of acting with approval. Nothing on the bench writes to a business system, and fully autonomous execution is shown only as a path the rules refuse.