In either a local AI or a hosted AI arrangements, I'm seeing lots of posts of people saying they have a need for agents 24/7, assuming they're in the software space. What are these agents actually doing? How close to the frontier are these models, and what does it cost?

I run a SaaS business, but I feel like agents aren't really working for me as once the last large task finishes that I assign for the day. I have a huge backlog that I would like to automate but I'm struggling to comprehend how people get quality outputs from tasks and where the user is supposed to be in the pipeline and how things go from idea to task to automated build to what I assume is the final human intervention which is the review process.

Top line, I'm assuming they're using some sort of issue tracking system like Linear or GitHub issues into either a cloud agent or a work tree where it's built out and tested? What about for things that aren't bug or issue based, but are actually features? How much context do you provide? How do you stop the AI just running with things instead of asking for clarity when required, etc.

Is there actually a net positive benefit to this setup, or do you end up just having to patch all of the work that was done autonomously overnight most of the time?