My experience with ChatGPT Agent so far: I've failed to find any use cases that _cannot_ be handled by Deep Research and yet can be successfully completed by Agent without running into any stumbling blocks like janky web forms or access restrictions. I'm sure I'll find some uses, but it will end up being a small fraction of tasks that come up in my workflows. If this is the case, it won't make sense to try to do new tasks using Agent unless it's a task that I would otherwise spend hours on (or would need to repeat on a daily basis). If my expectation is that Agent will succeed with a 5% probability, and it takes 10-20 minutes of trying painfully hard before giving up, it's not worth my time to even find out if Agent can do it. I would only use it if I somehow already knew that it's a task that Agent can handle. Given all this, I continue to think that task-specific agents will be more successful for the foreseeable future.

To be clear, this is as much a story about the usefulness of the Deep Research product category (which I do use on a regular basis) as it is about the limitations of the Agent / web browsing category. That said, most of the time I want DR to just gather data and produce a spreadsheet. The text it insists on producing on top of that function creates negative value for me by biasing my thinking. This is one reason I was excited to try Agent. But alas.

I found this example by @binarybits very interesting. This was one of the more successful things that he tried, and yet it is so far from reliable that it's not worth the trouble for the minor time saving.