Delegated, but Not Authorized: The question multi-agent systems have not been asking

Work in a multi-agent system crosses three kinds of boundaries: between the stages of one workflow, between different workflows inside one company, and between companies. Considerable effort is devoted to assessing the quality of multi-agent output, but far less to a question institutions learned long ago to ask before relying on the work: was this actor authorized to transfer this work to that actor? Multi-agent systems depend on delegation. Orchestration agents assign tasks, specialist agents produce work, consumer agents act on it, and work changes hands repeatedly with little human oversight. This paper makes the case that transfer authorization is distinct from access control and output quality. We identify five failure patterns that quality checks do not catch and five properties that safe transfer authorization systems need to address them. Authorization matters at all three boundaries, especially when agents in one company transfer work to agents in another. Our examples come from semiconductor design and manufacturing, although the problem applies across industries.