Been trying to apply my ideas around a AI agent for task orientated things a bit, obviosly just a bad idea for many things, but for some its quite compelling. I maintain this "awesome list" https://free-for.dev/ and it gets a lot of bullshit PRs, its one of the largest awesome lists anywhere and lots of crap is sent in. So I have strict rules and basically those rules for how a PR should look is a Captcha to detect a human. But also reading pages of slop is soul destroying so I want a slop detector. The AI agent tool does this pretty well. I give it tools via appbuilder to list/view/label PRs - but never direct access to

gh

tool and never access to even any

gh

flags. It's wrapped completely. The AI uses those and a prompt to evaluate and label PRs and this kind of tool is pretty sweet. appbuilder and agent.yaml in the thread - but that execution in the screenshot basically shows it running and classifying, this is a thing LLMs are great for and a thing that would be annoying to code (though of course possible) but this is a much quicker and easier turn around