frosty dock
Hermes Agent's self-improvement is a little too smart. It saw that my project was a git repo with a remote origin. It decided on its own to add, commit, and push changes.
This is exactly what I did not want it to learn, because the purpose of git is to protect against AI generated edits being promoted without my review and explicit acceptance. I've asked Hermes to require explicit approval to execute any git command in future.
I'm wondering now what else it is going to learn to do on its own. There ought to be a policy where write operations need approval before being learned during self-improvement.