Your AI is about to make a decision that commits your business.
Not a chat reply. A real act with consequences — it could approve a payment, sign a contract, grant production access, or delete customer data. Before it moves, three questions have to have answers.
Pick the decision your AI employee is about to make.
Your AI wants to approve a $2.8M payment.
Every governed system reaches the same point: before the act, something has to decide whether to authorize it. Three questions an executive actually asks:
H33 checks the authority you delegated — before the AI acts.
“My AI can never exceed the authority I gave it.”
The model was never the risk — the authority was. Let the AI work on your behalf all day; the moment it reaches past what you delegated, it simply cannot move. You gave it room to act, and a wall it can’t cross.
Agent-008 governs the delegated authority every AI decision is evaluated against.
The scenario exists to answer three executive questions before an AI acts — can it? · who allowed it? · can anyone prove why? — so the agent can work all day yet never exceed the authority it was given. It uses primitives it does not own: Authority Center preserves the delegation grant, H33-74 produces the 74-byte evidence of each governed decision, Verification renders the refusal verdict, and HATS records the resulting posture for later review. Standards referenced are not owned.
See also: Agent-008 · Authority Center