I’d Fire an Employee for This: Three Checks Before an AI Agent Acts
Jimmy Explains · Jimmy Standaert

Your agent has permission. Its pricing is out of date. Here’s the check I would make before it sends that quote. Imagine a customer asks for updated pricing. The agent can access the CRM, draft the quote, and send the email. All three permissions are valid. But the price sheet it used expired yesterday. That’s the problem behind Permission Is Not Proof. First, make the evidence visible. Which price sheet supports this quote? Who issued it? When was it valid? If it’s stale, missing, or contradicted, pause the consequential action and resolve the gap. Second, bind approval to the specific decision. The person reviewing the quote needs the evidence, uncertainty, recipient, attachment, and proposed terms. An approval should cover that package within a defined scope and time window. Now watch what changes. If the agent edits the amount, swaps the attachment, or changes the recipient after approval, the earlier approval may no longer apply. Our proposed contract requires the action to be checked again before execution. We call the reviewable record a Governed Decision Object. Think of it as the decision, its supporting evidence, its authority, and the exact intended action, kept together. Third, prove the effect. A successful tool response is useful evidence, but the receiving system needs to confirm what actually happened. Record the source-native receipt and read back the relevant state. If completion is uncertain, reconcile before retrying. Don’t turn a timeout into a duplicate send. The paper’s appendix makes these boundaries concrete with an illustrative decision object and six failure cases. Expired evidence, changed recipients, revoked authority, and missing receipts each have a defined expected response. Those are proposed acceptance criteria. They are not executed tests or production proof. This is the direction behind Whitewater AIOS: connecting decisions, evidence, authority, action, and receipts across the AI tools you use. LLM-agnostic. Model-agnostic. Agent-agnostic. The human belongs inside that decision process, with enough context to challenge, approve, narrow, or stop the action. Human decision integrity has to survive delegation. Here’s the practical rule: before an agent acts, make the decision reviewable. After it acts, verify the effect. Read Permission Is Not Proof in the PDF attached to my LinkedIn post. The compact technical appendix is included. Keep the decision visible. Bring back the receipt.