Large-scale enterprise systems increasingly rely on centralized multi-agent architectures, where a coordinator hands user requests to lightweight, domain-specialized sub-agents. This keeps the system modular and scalable, but it creates a reliability problem: a sub-agent may be asked to handle something outside its expertise. If it guesses anyway, the error can propagate through the whole pipeline.
The paper proposes EARS (Explanatory Abstention for Reliable Sub-Agent Modeling), which gives sub-agents an explicit "abstain" option. Instead of forcing a confident but potentially wrong answer, a sub-agent can decline and provide a short explanation of why it cannot complete the request. The coordinator then uses that explanation to decide whether to retry, reassign to another sub-agent, or escalate to a human.
The key contribution is making abstention actionable. A bare refusal would leave the coordinator blind, but an explanatory abstention turns a failure into useful signal. This matters especially at scale, where a coordinator cannot manually inspect every sub-agent output. The paper is a single arXiv preprint, so it has not yet been peer-reviewed, and no comparison with other abstention methods is available from the abstract alone. Still, the idea directly targets a known failure mode in delegated AI systems.