A consent receipt is useful only if the agent knows when it no longer applies. Stale consent checks compare the original approval evidence against the current task, state, recipient, account, and output surface before the agent acts again.
How-to workflow
Personal AI agents often resume work after the user has moved on. A stale consent check prevents an old yes from authorizing a new action when the evidence, account, recipient, browser state, amount, or public output has changed.
The stale check should read a structured receipt: requested action, evidence bundle, allowed scope, excluded scope, expiry rule, audit range, replay link, and rollback owner. Conversation history can help explain intent, but the receipt is what the agent should enforce.
For a text message AI assistant, this keeps approval lightweight while preserving enough context for future reuse decisions.
Browser page, draft, price, public URL, or selected account no longer matches approval evidence.
Recipient, collaborator, vendor, or customer context differs from the original receipt.
Approval window, launch context, travel window, or temporary exception has expired.
A computer-use cache can compare current page state against the screenshot or extraction the human approved.
For AI agent website-building, stale checks should compare page diff, source list, planned URL, and deploy target.
Scrubbed rule
If the current action cannot be explained by the original evidence, scope, expiry, and audit trail, the personal AI agent should ask again.
Checklist
The practical stale-consent gate is small. It should be fast enough to run before every consequential action.
Does current proof match what was approved?
Is the action still inside allowed bounds?
Is the approval still within time and context?
Did the previous run stay inside scope?
FAQ
The point is not to re-prompt constantly. The point is to prevent old approvals from leaking into new facts.
Yes, either by time, state, scope, recipient, amount, output target, or audit evidence. Permanent consent is usually too broad for personal agents.
Yes. A narrow valid receipt lets the agent proceed safely, while stale evidence triggers a focused reapproval instead of a vague prompt.
The agent should explain what changed, show the relevant evidence, and ask for a new scoped approval.
The NIST AI Risk Management Framework supports governance framing, and the OWASP Top 10 for LLM Applications supports unsafe-action risk framing.
Stale consent checks let personal AI agents move quickly without treating every previous yes as permanent permission.