When Agents Delegate: Accountability Chains in Multi-Agent AI Systems

Abstract

Ethical frameworks for the oversight of autonomous AI systems generally assume a dyadic relationship: one human principal authorizing and supervising one artificial agent. Deployed practice has moved on. Contemporary agentic systems decompose tasks and delegate them onward, to sub-agents they instantiate, to third-party agents discovered through interoperability protocols, and to tools whose own behavior is agentic. Each hop in such a delegation chain attenuates the three prerequisites on which accountability depends: authorized scope, attributable intent, and reconstructable provenance. This study analyzes the resulting accountability gap. We contribute a taxonomy of delegation structures in deployed agentic systems; an analysis of four failure modes by which delegation dissolves accountability, which we term authorization dilution, intent drift, provenance loss, and responsibility laundering; a normative framework, the Delegation Accountability Chain, comprising five conditions under which multiagent delegation remains compatible with meaningful human oversight; an audit of five widely used interoperability protocols and orchestration frameworks (the Model Context Protocol, the Agent2Agent protocol, LangGraph, AutoGen, and the OpenAI Agents SDK) against the five conditions, finding that none satisfies any condition chain-wide at a mandatory level; and a mapping of the conditions onto the human-oversight obligations of the EU Artificial Intelligence Act and the NIST AI Risk Management Framework, both drafted for the dyadic case. We argue that the question of where the buck stops has a defensible answer in multiagent systems only if delegation is constrained by design.