Why I stopped trusting GPT wrappers for real workflow tasks

Everyone keeps raving about these AI agents that book flights or write code, but I tried three of them in the last month and each one messed up a simple step. One from a startup in Austin doubled a meeting invite, another mixed up time zones on a project deadline, and the third gave me a refund total that was off by $45. The issue is people are treating these as finished products when they're really demos with a payment page. I'd rather use a basic API call and build my own checks than hand over control to something that can't even handle edge cases yet. Has anyone else had a similar letdown with a specific AI agent, or am I just picking the wrong ones?