The article argues that current AI incidents stem not from model capability or permission failures, but from a lack of procedural oversight. It contends that society has incorrectly granted AI the same trust it never gives humans, who must rely on records and checks before acting.
- The author proposes splitting responsibility among three parties: tool builders, users, and agent adopters.
- An agent should verify inputs via a checklist, source validation, unknown handling, and record-keeping.
- The LLM's role is limited to reasoning, while the agent controls execution based on declared grounds.
- Current systems lack fields to declare conditions, forcing reliance on the model's inference which cannot be audited.
The piece concludes that stronger models require stricter separation of concerns, shifting from "running unless forbidden" to "not running until confirmed" by a verifiable record.