KorvaJoin the waitlist
Tech

Make a Different AI Check the First AI's Work

Shopify and RunLayer treat AI results as unproven claims until a separate agent sees evidence.

AI image
The AI Daily Brief: Artificial Intelligence News and Analysis · 7 Things AI-Native Companies Do Differently

Shopify does not take its AI's word for it. No security alert counts on its own.

The proof has to be practical. The agent must build a demonstration that shows the flaw can actually be triggered. If it cannot, the warning is thrown out or given lower priority.

The check is also kept independent. A different model from the one that found the problem reviews it, so one system is not grading its own work.

RunLayer, another company using AI agents heavily, follows the same pattern for software fixes. The agent that makes a fix must submit its proposed change with visual proof of the before and after, such as images or a short video tour.

Then a second agent reviews that result as an adversary, looking for weaknesses. Even after the change is released, a quality assurance agent attempts to defeat it during live use.

The idea is that the bottleneck has moved from producing work to verifying it. The job is not done when the first agent produces an answer, and verification needs its own step and its own system.

To try it this week, pick one repeated AI task and add a separate checking pass. Use a different agent for the check, demand evidence like a test, screenshot, or walkthrough, and reject anything that cannot prove itself.

Questions people ask

Why does it matter that a different model does the checking?

Using a different model reduces the chance that the checker shares the same blind spot as the first agent. It is the same logic as getting a second opinion.

What counts as proof if I am not writing software?

Good proof is something you can observe, not just an explanation. In software that can be a test that triggers the bug, screenshots of before and after, or a short recording. In other work it might be a source document, a calculation, or a checklist of required facts.

Should I still have a person review the work?

Yes, but the check has to fit the risk. For low stakes drafts, a second AI pass for facts, numbers, and missing steps may be enough. For money, safety, legal language, or anything published, keep a person accountable and have that person review the evidence.

More in Tech