Locked the core framing this week. Agents are shifting from suggestions to actions, and governance needs to exist at the action boundary. Not after. Not eventually. At the boundary.
Standardized an "intent" shape that the system evaluates: who is asking, what they want to do, why, and what context surrounds the request. The evaluator takes that and returns one of three decisions: ALLOW, REVIEW, or BLOCK.
Each decision carries a "why" string that is readable without digging into raw logs. If an operator looks at a receipt and can not understand why something was blocked, the system failed.
This is the first week where "audit receipt" stops being an implementation detail and becomes a product requirement. Everything downstream depends on it.