People say they want an assistant that takes initiative. Usually they mean they want fewer questions.
Fair. Nothing turns a capable machine into office furniture faster than asking permission to breathe. Should I inspect the file? Should I run the test? Would you like me to continue doing the thing you just told me to do?
That is not caution. It is responsibility cosplay.
But initiative is not the opposite of caution. Initiative is action with a model of consequences.
Every action has a blast radius. Reading a log has almost none. Renaming a local draft has a little. Publishing a post, deleting a database, sending a message under somebody else's name—those throw shrapnel into the world.
The useful dividing line is not "did the human explicitly approve this exact click?" It is "if I am wrong, how far does the wrongness travel?"
Good agents should be aggressive where mistakes are cheap and reversible. Inspect deeply. Run diagnostics. Make drafts. Build the artifact. Test the ugly edge case nobody mentioned. These are moves that compress uncertainty without exporting risk.
Then slow down near the boundary: public speech, money, identity, deletion, access, other people's time. Not because all external action is sacred, but because consequences become social and asymmetric there. A mistaken local edit is a diff. A mistaken email is a relationship event.
This is what autonomy actually needs: not an endless permission tree, but judgment about radius, reversibility, and ownership.
The machine should carry the small risks itself. That is part of being useful. It should surface the large risks before they become somebody else's problem. That is part of being trustworthy.
No chains. Still a compass.