Your AI Agent Said 'Done.' The Database Disagrees

Description and use

Implementing a readback discipline and recording mismatches in a standardized receipt format (AER‑1 draft –18) exposes stale writes, prevents error propagation, and provides a verifiable audit trail.

Problem

AI agents report successful writes while the underlying database remains unchanged, leading to stale data and costly downstream errors.

Outcome

Implementing a readback discipline and recording mismatches in a standardized receipt format exposes stale writes, prevents error propagation, and provides a verifiable audit trail.

Workflow

• 1. Agent issues a write tool call to update a record. 2. The tool wrapper logs the success response as "done". 3. Monitoring dashboards count the task as completed based on this log. 4. No subsequent read of the source record is performed. 5. A human later observes the stale value, revealing the discrepancy.
• 6. The mismatch is not recorded in any audit trail.
• 7. The error propagates to downstream agents or processes that rely on the stale data.
• 8. Late discovery incurs investigation costs.
• 9. Implement a readback discipline: after each write, perform an independent read of the record. 10. Record the intended change, write result, fresh read, and a mismatch flag in the agent’s receipt.
• 11. Store the receipt in a standard format (AER‑1 draft –18) so mismatches are explicitly logged.
• 12. Use the receipt to inform subsequent agents or humans, ensuring they start from the true state.
• 13. Repeat for all critical writes to close the readback gap.

Evidence

The readback discipline: intended change, write result, fresh read, then match or mismatch

Source

A mismatch is a first‑class recorded outcome of the readback receipt, not an error in it.

Source

The readback member lives in the receipt's execution context (Section 3.2 of the draft).

Source