The architecture: six layers and a control plane

Each layer has one job. The control plane applies to every one of them.

The bench follows the Ardinois AI and Agentic Capability Architecture. Its principles are simple: use the least specialised design that meets the need, keep surface and runtime separable, route every model call through a control point, bound authority twice (by who acts and by which door is used) and treat anything read from outside as data, never as instructions.

The six layers

1. Interaction

Where people meet the capability. The chat workspace, Copilot Studio and Foundry are all surfaces at this layer.

2. Orchestration

The loop: gather context, decide, validate, act, record. Stable steps stay deterministic, and a model handles only bounded tasks inside them.

3. Models

The reasoning models, reached only through a gateway that attributes every call to a use case and an owner.

4. Tools and actions

What the runtime can do, through an allow-list of tools with declared contracts and versions. Reads and writes go through approved doors only.

5. Knowledge and rules

What the runtime knows, from plain instructions up to a formal semantic model, using the simplest form that serves the use case. See the knowledge argument.

6. State and evidence

What is remembered. Working state serves the run and is cleared. Records and audit evidence go to governed stores with their own retention.

The control plane

Four concerns cut across all six layers: identity (who acts, and with what rights), trace (one correlation identifier recorded by every component that touches an action), cost (usage measured and attributed, with scale to zero by default) and evaluation and release (tested before release, and tested again when something changes).

Four routes and four release levels

The architecture sorts capabilities into four routes, from productivity assistance (route A) through fixed workflows (B) and bounded agents (C) to autonomous agents (D). Release has four matching levels, from registration to formal governance release. The more freedom a capability has, the more evidence it needs. On the bench, route D is not run: it appears only as a refused path, together with the rule that refuses it.