The YOLO mode problem. How do you handle AI agent permissions?
Real question for anyone running coding agents daily.
I started counting how often Claude Code and Codex ask for permission during a typical session. It's around 100 prompts per hour. Read file? Yes. Run test? Yes. Lint? Yes. git push --force? Also yes if you're zoned out.
The industry gives you two modes right now:
1. Approve everything manually. Safe, but you're handcuffed to the terminal.
2. Skip all permissions. Fast, until rm -rf ./src happens and now you're having a different kind of day.
No agent has per-tool rules built in. No "auto-approve reads, escalate deletes."
How are you dealing with this?
- Approving everything manually?
- Running YOLO mode and trusting the vibes?
- Built custom scripts?
- Just avoiding dangerous tools entirely?
I'm building something in this space and trying to figure out if per-tool permission rules are what people actually want, or if most devs are fine with the binary choice.
Replies
Approving everything manually is my way to go and the only way for me to feel safe and in control of what is going on
I keep YOLO mode on by default.
The way I make it work is by setting rules: enough rules, and good enough rules. Add constraints.
That might sound like the opposite of the essence of YOLO, but my counterintuitive experience is that the better governed the AI is, the faster the work gets.
It spends less time at unnecessary decision points, and it prevents the project from wasting time on avoidable fixing and repair.
And of course: always use git, so I can roll back when needed.
I don't think the real problem is permissions. It's confidence.
If an agent asks for permission every few seconds, you stop thinking and start clicking "Approve." That's not safer, it's just slower.
I'd rather have a system that asks less often but explains why a high-risk action needs approval. Fewer interruptions, better judgment, and a much lower chance of approving something dangerous out of habit.