My research agent fabricated a product feature and i almost pitched it to my vp

I work in logistics planning. A couple weeks back my team was evaluating a freight forwarding platform for a sales review, and I had my research agent summarize the provider's EU compliance features. The summary it gave back looked solid. One feature jumped out, something like a "dynamic customs filing router" that supposedly automated tax declarations on the fly. I pasted it into the deck without checking. During the review, our VP stops on that slide. He had just left that shipping company the year before. He said the feature was never built, it was a brainstorm idea from 2023 that got killed because of customs rules. I wanted to melt into the floor. I spent the rest of the day sending correction emails and reworking the deck. The part that still bugs me is that the agent wasn't even guessing wrong in an obvious way. It was wrong in a very specific, very believable way. And when I asked it to verify, it just doubled down because it was reading its own previous output. Same model, same prompt window, same blind spots. I have been rebuilding the workflow since then. The rule I landed on is simple: the thing that writes the answer cannot also be the thing that checks it. If you want a real check, it has to be a separate run, ideally with separate context, separate search, and no memory of the first answer. A friend mentioned a new agent verification tool called Apodex that is built around this split, but honestly you could do the same thing yourself with two different model calls and some web search. The embarrassment alone is enough to make me verify everything manually from now on.