My agent booked the wrong flight, confidently. The receipt is a lesson in delegation

The setup felt careful. I wrote out my preferences like a policy document: window seat, no red-eyes, prefer nonstop under three hours, budget cap per trip, loyalty program above all. I gave the agent access to a dedicated travel card with a low limit. Then I let it book my October trip while I watched something else. Every rule was followed. The trip was still wrong.

What it booked: a nonstop, window seat, under budget, on my preferred airline. Also: a departure day one day before the event I'd told it about in a separate conversation — because I'd mentioned 'the 14th' in one thread and 'the conference' in another, and the agent resolved the conflict the way a literal-minded temp would. The flight was perfect for a trip I was no longer taking.

Here's what bothered me, and it isn't the mistake. When I've made this error myself, there's a moment of doubt — something feels off as I click buy. The agent felt no such friction. It executed the resolution of an ambiguous instruction with the same confidence as the unambiguous ones. Confidence is not a signal of correctness; it's a rendering choice.

The fix wasn't smarter rules. I tried writing more rules and got a document the agent followed even more literally into new corners. What worked was a checkpoint: the agent now drafts, I approve, then it books. One click of human judgment at the point of no return. The dream of full automation dies at the same place most process design dies — at the handoff.

I still use the agent. It saves real time on the 90% of travel booking that's genuinely mechanical. But I've stopped asking whether agents are good enough to trust and started asking which decisions deserve a checkpoint. The answer for flights, apparently, is one. The answer for your use case is probably one too.