Bots draft, humans approve
Agents
Adapted from a thread on XView on X
My rule is simple: bots draft, a human approves every send.
A smarter model isn't the fix
If a model can one-shot people, the fix isn't a smarter model. It's a person in the loop who reads before anything ships.
The approval gate is the whole thing
Fixing a bad scene plan costs a few tokens. Fixing it after the subagents render costs real credits. The gate sits where mistakes are still cheap.
Ownership is the hard part
One commander per project is right, but the commander isn't the hard part. Ownership is.
What made my bot stack work was rules like these:
- Only one agent can move a task.
- Nothing goes out without a yes.
Without that, autonomous just means agents undoing each other at 3am.
Same rule, every time
They draft. I approve before anything ships.
