I'm using an llm to read, find, and categorize text.
Things were going well with qwen3.8, but I was struggling to get it to understand numbers and context properly. I tried adjusting it with Astra, burning through tokens in the process, but I kept failing. It would seem to be getting better, only to regress back to its original state. It was disheartening.
Then I realized that Astra was controlling the qwen model directly through Python to get answers. In other words, because I was trying to process things with one-off calls, I was getting different answers every time I called... Ugh... This had been going on for weeks, and I didn't even realize it.
So I wrapped it in a simple harness called pi and told it to do its thing....
And wouldn't you know it, after two weeks of frustration, everything worked perfectly!!! I was so upset that I had wasted all those tokens, but seeing it work made me happy. I didn't say anything to Astra about it.
Then I started wondering, if even the weakest pi could do this well, how much better would other options be? Things are working so well right now that I can't bring myself to touch anything according to the developer's law.
I'm curious what other people in the AI community are using for local llm coding or agent harnesses.
It seems like AI performance is half model, half harness, so I'm interested in seeing what other harnesses are out there.
