>>105928948 (OP)I was working on a project and decided to put github copilot in agent mode using claude sonnet 4. Gave it access to the whole documentation of protocol I was trying to use and told it to analyze my code base and the docs in order to figure out a solution to a very specifically described problem statement.
It got to work and I was amazed. It read the relevant docs, searched for information, came up with an architecture, implemented it, evaluated its own work, wrote tests and interpreted the results and iterated like this for a while.
But in the end it came up with a ridiculously overcomplicated solution that didn't even work, and that I had to spend exponentially more time on trying to fix than it "saved" me from doing the work myself.
In the end I wrote my own solution that was way more elegant and not even 10% of the lines of code shat out by copilot.
AI may be able to write "code" but it can't write programs. Just like any retard can write words but writing a book is a whole different ballgame.