Many users agree explicit rules, sandboxes and approvals should be default for AI coding agents to avoid yolo risks like deleting Mac files, while others argue reducing blast radius matters more than model-based checks.
Based on 5 visible X reactions from 14 accounts; directional sample.
Ask a question below.
Published answers will appear here.
The key protection is reducing blast radius, not asking one model instance to approve another. Both agents may share the same mistaken assumption, and a perfectly “reasonable” command can still be disastrous in the wrong directory. Run the agent in a disposable worktree or container, mount only the repo it needs, keep $HOME and credentials out of reach, and checkpoint before every destructive step. Then approval becomes the last layer, not the only layer. A safe coding agent should be allowed to make mistakes. It just shouldn’t be allowed to make irreversible ones.
@reach_vb this is right, and it goes past the terminal. an agent in your browser is the same yolo risk with a bigger blast radius: your logged-in sessions and accounts. same rule: watch what it does, gate the irreversible stuff. visibility isnt friction, its the point.
@reach_vb this should be default advice. Approve for me + rules against destructive shell commands saved me more than once. yolo is fun until the agent finds rm
@reach_vb rules docs are the missing layer for most agent setups. sandbox + explicit allow/deny beats hoping the model is careful
@reach_vb Good advice there, thanks. I think we’ve all got a bit too comfortable yolo-ing 😂
PSA: if you care about your data and infrastructure then please *don’t* run your coding agent in yolo mode To secure your workspace I recommend the following: 1. Use “Approve for me” - you can have codex review what it’ll run with a sub agent to make sure that it’s not destructive! 2. Use Rules - you can set up a rule for codes to never run destructive commands like “rm - rf”, this needs to be in a.rules file and you can ask codex to create it too 3. Setup a PreToolUse Hook - you can ask codex to set this up for you as well
Rules https://learn.chatgpt.com/docs/agent-configuration/rules
Auto Review https://learn.chatgpt.com/docs/sandboxing/auto-review
Hooks https://learn.chatgpt.com/docs/hooks
Many users agree explicit rules, sandboxes and approvals should be default for AI coding agents to avoid yolo risks like deleting Mac files, while others argue reducing blast radius matters more than model-based checks.
Based on 5 visible X reactions from 14 accounts; directional sample.
Ask a question below.
Published answers will appear here.
PSA: if you care about your data and infrastructure then please *don’t* run your coding agent in yolo mode To secure your workspace I recommend the following: 1. Use “Approve for me” - you can have codex review what it’ll run with a sub agent to make sure that it’s not destructive! 2. Use Rules - you can set up a rule for codes to never run destructive commands like “rm - rf”, this needs to be in a.rules file and you can ask codex to create it too 3. Setup a PreToolUse Hook - you can ask codex to set this up for you as well