Ask HN: Why are AI code editors not continuously working?
I am getting more and more baffled by the incompetency of AI code editor developers. Why not simply have a "janitor" always running in the background, doing exactly the right refactorings, ensuring clean, well-tested code etc.? It would help so much against the AI slop problem. And it would be so simple to implement: Just some intelligent prompting or even static and dynamic analysis to detect re-occurring patterns in code and execution paths.
6 comments
There is a tension between provider and user.
Like a gym membership the provider wants you to pay a monthly fee and not use it 24/7.
That’s one of the reasons Claude Code has short OAuth renew times. If you run it in a loop, it eventually asks you to login - and it is not time based as far as I can tell.
On PAYG you’d be paying for endless “nothing to do” messages.
But as others have pointed out, the end user is the AI code editor developer. As you might have realised - I do have “always running in the background” agents.
But mine use GitHub (and our local hosted Gogs) actions on the issue trackers to run the agents on demand rather than continuously crawling the code.
I use labels to co-ordinate add-feature, add-documents, fix-bug etc. to determine which agent takes a look and have custom prompts for each label.
Like a gym membership the provider wants you to pay a monthly fee and not use it 24/7.
That’s one of the reasons Claude Code has short OAuth renew times. If you run it in a loop, it eventually asks you to login - and it is not time based as far as I can tell.
On PAYG you’d be paying for endless “nothing to do” messages.
But as others have pointed out, the end user is the AI code editor developer. As you might have realised - I do have “always running in the background” agents.
But mine use GitHub (and our local hosted Gogs) actions on the issue trackers to run the agents on demand rather than continuously crawling the code.
I use labels to co-ordinate add-feature, add-documents, fix-bug etc. to determine which agent takes a look and have custom prompts for each label.
It's not the fault of the "code editor developers." The underlying technology, GenAI, is non-deterministic. When you add input that varies widely in quality (prompts, existing code) what you describe is actually quite difficult.
> And it would be so simple to implement: Just some intelligent prompting or even static and dynamic analysis to detect re-occurring patterns in code and execution paths.
So just do it, then? Or are you just as incompentent yourself?
So just do it, then? Or are you just as incompentent yourself?
> doing exactly the right refactorings
Are you aware of any LLM models that can do that?
Are you aware of any LLM models that can do that?