TUIs have this really nice attribute of running local to the data you care about (web apps have a lot of complicated glue code required to do interesting things), and in a UI that is incredibly simple to test and iterate on because the output can be serialized to a text file with very minimal loss of fidelity.

It is hard to remove the rise of agentic flows and LLMs generally from this discussion. The fact that LLMs understand text in a first-class way and have had more time to get good at them, means that they can iterate and build apps and features that might bog down a model in the context of a GUI app and all of the additional context required for the larger framework and deployment and testing story.

These attributes and the current environment with LLMs mean that it's really easy to build ephemeral UIs that serve a purpose right now, and then consider productionizing or generalizing at a later date. If you decide to do this, you have the really high fidelity tests that make it easier.