Currently, agent applications are in the middle of the turbulent period of rapid LLM development.

The successful business models are all integrated intermediary + blind box drawing services:
Providing a layer of wrapping over bare agent software (CLI, Codex, etc.),
significantly improving the user experience in specific vertical fields by ten times, simplifying the drawing process, stimulating urgency, and guiding users to draw the next piece of code, image, or video through various methods.
In fact, these applications are still selling traditional SaaS services,
and they might even survive until the AI bubble bursts, with AGI never arriving,
and the demand for everyone to customize their own software (which was always a pseudo-demand and a concept) will never materialize.