Beyond agentic coding

I find chatbox-style applications annoying, but it might just be because I'm associating it with those online helpdesk type of things. And at work, where we had to use ChatGPT, I find agents to be...really eager?

Anyhow on a related note, I don't mainly use Emacs, but one of the more interesting things I've seen is gptel. It's leaning into the fact that everything is a text buffer in Emacs, which makes glueing stuff together really amazing, and it retains plaintext/history and data because it's an org or markdown file! For me, it retains the calmness.

I suppose one of the addicting things of LLMs is the fact that one can easily glue things (unsafely) together, malleability of code and services.

Concepts of Plan9, Emacs, arcan or Glamorous Toolkit might be easier to sell to management with such a pitch, might have to switch calm for productive.

I appreciate a wild take. I read the first sentence as "I'm generally pretty pro-AI with one major exception for the one thing it's good for."

I do like the examples of calm tech powered by LLMs — semantic file browsing, automatic commit reorg (I recently drafted a Claude skill to do this), and file lenses — but in my experience of using agentic coding CLIs, the chat interface is a pretty fine way to accomplish all of these things. I do want to see more real UIs like this (I am mostly a UI dev after all) but I think in order to work they'll have to be pretty dynamic (read: LLM-generated on the fly) since model capability and scope of work is such a moving target right now.

Another thing I'd note is that all the arguments presented against agentic coding here — they demand attention, they stand between us and the code, they undermine calm — all apply to typecheckers. Is that supposed to mean typecheckers are also bad, and if not, why not?

Ironically, the way I often am most productive with coding is by talking through the problem with someone.

If it performed better, it would feel like pair programming...