Have a second AI audit the first one: check what the customer receives, not whether the job ran

An automated content and sales system reported tasks as done when the output a person would see was broken, and corrections were saved without confirming the mistakes stopped.

A carousel job reported success though the images were unreadable at phone size. The author had a separate AI with no history of the setup compare, for each part of the system, the instruction, the code, the latest output and the end result a person would see. A second model reviewed the repairs and the author decided what shipped. Findings included success measured by whether a job ran, cluttered memory, conflicting instructions between files, feedback saved without checking it worked, and no one following a real buyer through checkout, email and download together.

The audit found problems and repairs were made, but the post gives no counts of defects found or fixed and no before and after metrics. The scope audited was eight directors, 44 smaller agents, 74 skills and 56 scheduled jobs.

The specific repairs and their results are not described. The post notes that two models can agree and still be wrong, since one wrongly applied a rule from another platform and withdrew it, so the audit informs human judgment rather than replacing it. The auditing model's vendor and version are not stated, and the post offers a prompt as a reader takeaway.

An existing AI-run workflow, access to a second AI model, and time to follow a real customer path end to end.

• Belonging & Bots (Substack) ↗ Firsthand write-up, published September 15, 2026

Source published: September 15, 2026. Last reviewed here: October 11, 2026. Spot a mistake? Tell us.