Why doesn't the protocol have any code? #24
-
|
It seems we're depending entirely on the agent to decide what task it's supposed to be doing right now. I think I more or less understand the idea that the content of the markdown files adhere to a schema that unambiguously determines what the next task should be. But if the agent doesn't read the files, or if it has an older copy in context, it could hallucinate and do the wrong task. Why don't we have code that's like find_next_task.py that scans the markdown files looking for unchecked boxes? And probably it should also be checking the git status because sometimes git commit or github pull request would block moving on to the next task. It can then point to the correct snippet of the protocol that describes the current task so that if the agent has for some reason neglected to read the protocol it will have another chance to do so. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Simply because that was the easiest thing to implement :-). I'm open to this approach if you're willing to implement it and see if it works. Would definitely consider that PR. My gut, however, is telling me it might not work all that well. Because now you're replacing one problem with another problem: ensuring the agent knows to run find_next_task.py at the right time :-). If it knew that, it would have just done the next thing itself. Give it a try, let me know how it goes! |
Beta Was this translation helpful? Give feedback.
Simply because that was the easiest thing to implement :-).
I'm open to this approach if you're willing to implement it and see if it works. Would definitely consider that PR.
My gut, however, is telling me it might not work all that well. Because now you're replacing one problem with another problem: ensuring the agent knows to run find_next_task.py at the right time :-). If it knew that, it would have just done the next thing itself.
Give it a try, let me know how it goes!