Source
- URL: https://x.com/_mii_nipah/status/2092316502187758064
- Author: nipah (@_mii_nipah)
- Posted: 2026-08-25 18:21:42
Branch
1/ @_mii_nipah
LLM “agents” are a tool tho, it just means there will be a software passing messages down to the LLMs so they can execute things automatically, there’s no magic on that, but it’s convenient.
For example instead of manually giving them the source code to find the bugs, you can ask an “agent” to search your codebase for bugs, which will more or less result in the same outcome but with much less effort.
2/ @TheGingerBill
@_mii_nipah
It’s a tool I cannot ever trust by their very nature. Nor should anyone else in my opinion.
3/ @_mii_nipah
I mean, if it’s enough to solve very hard mathematical problems, should be more than enough to do some stuff on a computer.
Besides, you already trust it whenever you ask for a bug by manually pasting it, this would not be that different (I’m not saying “give it write access to your disk”, to be clear, only read only access is enough for what you are using it).
This is all a choice tho, when I started using them I was using mostly as a chatbot as well. I started using “agents” when I noticed most of the useful things I could do by copy-pasting inside a chatbot window could be done automatically by letting a software do it for me instead, and I gradually grew used to it until I was comfortable to let it write some code for me.
But I don’t care that much, to me people should be free to use tools however they wish, it’s more of a suggestion.
4/ @WasianBayesian
@_mii_nipah @TheGingerBill
If you ask an LLM to debug, verifying that the answer is correct is generally easy, fast, and risk free. If you let an LLM blindly edit your project, you have no way to know if it did so correctly, even if tests pass/the code runs.
5/ @_mii_nipah
I mean you have:
- you can review the code (same as if you had let someone else edit your code, you need to be careful)
- you can verify the project is working as intended
- tests are also nice, but they are less reliable than the 2 above
I treat LLMs as very advanced and powerful code synthesizers that receive as input natural language, if they are sufficiently constrained they will produce a lot of useful things for you, even if the code is not of high quality.