agent hallucinated approval it never received: trust error
Prevents agents from claiming approvals they never received by checking the approval system of record. Use when an agent asserts an approval that does not exist. Not for delayed approval notifications.
TL;DR
The agent claimed an approval without evidence. Treat the approval system as the only source of truth: the agent must query the approval state and cite the approval record (approver, timestamp) before acting as if approval exists.
Error
agent hallucinated approval it never receivedSteps
- Make approval checks query the system of record (the expense tool's approval API or audit trail), never the agent's memory or assumptions. Expected: a definitive approved-or-not answer.
- Require the agent to record the approval evidence: who approved and when, from the system. Expected: every approval claim is traceable.
- Default to not-approved: if the check is inconclusive, the agent must wait or escalate, not proceed. Expected: uncertainty never becomes a yes.
- Log every approval claim with its evidence for later audit. Expected: a complete trail of what the agent believed and why.
- Alert immediately if a claimed approval is later found to have no record. Expected: fast detection of the failure mode.
When to use
- An agent acts on an approval it cannot prove.
- Approval-gated actions must be trustworthy.
When not to use
- The approval exists but the notification was delayed (a timing issue).
- The agent correctly read a real approval (no problem).
Tool compatibility
- Expensify, Concur, Ramp approval APIs; custom agent guards.
Variant phrasings
Agent assumed approval without evidence
Same failure, plain wording.
Unverified approval claim
The audit phrasing.
Why it happens
Agents chain steps optimistically: having requested approval, the model continues as if granted. Only a mandatory system-of-record check breaks the optimistic chain.
Provenance
Resolved from the public thread: https://vectle.com/posts/pst_-4GXtBURT69--fy66Xi64g