I draw a huge amount of joy by working WITH my AI agents as opposed to them doing everything autonomously. However all the SOTA models seem to be actively doing the opposite: hide what they are busy with and take everything to completion with as litle interaction with a human as possible. Which models do you guys think are the best for pair-programming? And is there a way to change this behavior of the SOTA models, or are they just fundamentally designed to be like that?
Ask HN: Which model is best for pair-programming?
1–4 of 4 posts
Re: Ask HN: Which model is best for pair-programming?
#2I think this has more to do with the harness than the model.
Re: Ask HN: Which model is best for pair-programming?
#3I think this has more to do with the harness than the model.
You mean like the app that "controls" or "directs" it? The thing that executes the tools, etc?
Re: Ask HN: Which model is best for pair-programming?
#4I think this has more to do with the harness than the model.
You mean like the app that "controls" or "directs" it? The thing that executes the tools, etc?
Correct. There's no generic right answer and you didn't share any information about your own practices and preferences. Just mix and match until you find a combination you like. Most of teh harnesses are just simple user interfaces with combinations of sub-prompts, just like all the 'skills' that were the flavor of the moment a few months back are just prompts about how to do one specific thing. Some were written by people who are great at writing instructional text, many were just synthesized and by and LLM and marketed as some brilliant new innovation. The industry is awash in bullshit artists so stop trying to optimize and just use trial and error instead.