You cannot design the mistakes out of a language model. You can only decide how expensive each one is to the person on the other side. That reframing changes what you build.
Cheap to notice
An agent that quietly does the wrong thing is worse than one that loudly does nothing. Show the plan before acting. Echo back what it understood. Make the reasoning skimmable, so a wrong turn is caught in a glance instead of three steps later.
Trust isn’t built by being right. It’s built by being correctable.
Cheap to undo
Every destructive action wants a seam — a preview, a confirm, an undo. The goal isn’t to slow the agent down; it’s to make recovery so cheap that being wrong stops being scary. When the cost of a mistake approaches zero, the whole system feels calmer, and people let it do more.
Design for the wrong answer, and the right ones take care of themselves.