The system first parses a paper and identifies its academic discipline. It then dynamically routes the paper to the corresponding domain expert agent. Different disciplines, such as finance, mathematics, and law, are represented by specialized agents with their own domain knowledge, evaluation criteria, and review priorities. When a paper contains interdisciplinary content or requires multiple professional perspectives, several expert agents can participate collaboratively. A chief-review agent then aggregates, calibrates, and resolves potential conflicts among their evaluations before producing the final assessment.
The talk will focus on practical engineering challenges behind such a Multi-Agent system: how domain expert agents are designed and routed, how discipline-specific knowledge and evaluation rules are injected, how long-document context is managed, how multiple agents collaborate, how structured and stable model outputs are maintained, and how to evaluate whether a multi-agent approach actually improves review quality compared with a single general-purpose model. Through this real-world project, the audience will gain reusable approaches to expert-agent design, dynamic routing, workflow orchestration, and evaluation for other Python + LLM applications.