[seba_dos1]

Agentic harnesses go in the exact opposite direction to what I'd want to get from LLMs. I don't want another black box to (poorly) work on a black box for me, I want to be better at reaching into and understanding boxes that I already have in front of me. I don't want tools to autocompact contexts and store generated memories to facilitate long runs I have barely any control over, I want tools that allow me to painlessly craft a more relevant context for short ones. I don't want agents to author commits, I want them to use Git (or other tools) to get the information that I'm looking for when it's tedious to do it myself. I don't need them to do the fun and beneficial part of the job for me, I want them to do the boring parts that I already know how to do which block me from proceeding because my brain just isn't interested. Some of those things you can script yourself relatively easily, but the current tooling for LLM coding is absolutely atrocious and disconnected from programmer's needs.

The main output of my work is gaining a better mental model of systems I work with. That's what lets me grow and that's what makes people want to pay me rather than someone else to work on these things. Anything else, including the produced code itself, is secondary to that. In general I find it pretty hard, although not impossible, to use LLMs in a way that doesn't diminish my output, especially with this tooling that seems explicitly designed to make it hard. After all, reviewing things is so much harder than writing them yourself, and you can't feel accomplished by something you haven't done.