Thinking the same. I don't want Github approval process to sit in between me and the changes - the killer feature of claude code is being able to head it off as it starts to go down a bad path, and to code myself in between its steps.
Do you let juniors complete full features without asking questions or make them check in when they get flustered?
I do want to try out some background agents, but from my experience with Cursor’s (frontier model agents) frequency of going off the rails despite having rules and context to help avoid producing slop, I can’t see background agents being that generally useful yet.
Do you let juniors complete full features without asking questions or make them check in when they get flustered?