Agentic Postars

I'm turning postars.co into an agentic product. It has one clear advantage: I would be able to perform bulk operations.

That made me wonder whether it would have been better to build an agentic app from the start, without first going through a traditional interface with clickable buttons.

Actually, probably not. For an unverifiable task (to use Andrew Ng’s term), such as writing a post for LinkedIn or X, human evaluation of the output is still important, as well as experimentation.

For both evaluation and experimentation, we need an interface that lets us view the result and compare different versions. A chat window is not well suited to that. In addition, chat history is not structured, which means extra work to extract and format results for storage.

So the agentic layer should sit on top of a traditional interface layer. The two should work together (as least for unverifiable task).