Agent Mode in GPT-5: Upgrade or Downgrade?
The creator tests OpenAI's new GPT-5 Agent Mode on real-world tasks: finding AI trends and drafting documents, signing up for newsletters, and booking restaurant tables.
Agent Mode completes searches in 30 seconds and handles mid-task pivots smoothly, completing a brunch comparison in under 30 seconds compared to GPT-4o's failed 8-minute attempt.
However, the system struggles with email automation, hallucinating capabilities it lacks, and occasionally wandering into unexpected website versions.
The verdict: clear progress in speed and flexibility over GPT-4o, but the system remains buggy and confined—improvements without magic.
So, I tried GBD5's agent mode, and the results were complicated. I started by asking it to find the top five AI search trends of 2025, drop it in a doc, and email it to me. It found all trends in 30 seconds. So, aentic search, multimodal, voice native, content, providence, and structured data. But when it came to email, total stop. It just handed me a draft and said, "You do it. Thanks, I guess." And it even hallucinated, claiming it could convert to Google Docs, except it can't even log into Google at all. Then I told it to sign me up for Mary Hayne's AI and search newsletter. It started with, "I can't submit forms, but it did it anyway." And 2 minutes later, I was subscribed. That's a clear but messy win. Probably a bug or maybe even a hidden capability. Let me know what you think in the comments. In my final test, I asked it to book a table at G and Carlos. Then halfway through, I changed my mind and asked it to compare Brunch Bots instead. Agent struggled at first, wandering on to the website's Spanish version for some reason until circling back to Open Table, but the pivot itself
was smooth. In under 30 seconds, it gave me a comparison table of brunch options with directions ready to go. That's a big upgrade from GPT40 where it wasted 8 minutes and still failed at a simpler booking. So, is GPD5's agent mode better? Oh, of course. It's slicker. It's more flexible. There's no freezing. And those midtask pivots are just smooth. But don't expect magic. It's still fun in buggy. That's it for OpenAI's new agent. What tool should I test next? Drop it below. Oh, and don't forget to like and subscribe if you want to get more honest reviews like this