LLMs as Agents, LLMs at the Service of Agents, or Agents at the Service of LLMs?

In a panel entitled “Agents vs LLMs” organized during the recent annual meeting of the Italian agents and multiagent systems community, the “From Objects to Agents (WOA) workshop”, some intriguing and provocative questions on agents, agentic AI, and LLMs were made to the panelists. While the answers were often aligned, or converged towards an alignment after some interactions and iterations, the “emotional” attitude of the panelists towards Agentic AI turned out to be very different. Indeed, some researchers feel very frustrated by the fact that ideas that were born in the MAS community, are proposed as brand new ones from the Agentic AI community without any acknowledgment. Others see this as a confirmation that the ideas produced in the last decades by the MAS community were actually good. Consider for instance the case of object-oriented programming: it emerged in the programming language community in the 60s, it became the mainstream paradigm since the 90s, and nowadays it is so deep and pervasive in the computer science culture that no author even feels the need to cite something when mentioning OOP in a paper. It means that the idea was good. A similar thing may be happening to the AOSE and EMAS ideas. In this extended abstract, we comment on the “Agents vs LLMs” WOA panel and on what emerged from it, and we devise PlanchBDI, a practical proposal to take the best of the new – subsymbolic, opaque, effective, widespread generative AI – and the old – symbolic, transparent, “for connoisseurs only” – Belief-Desire-Intentions (BDI) approaches.

evento origine

 EUMAS 2025

progetto finanziatore

 ENGINES — ENGineering INtElligent Systems around intelligent agent technologies (28/09/2023–27/02/2026)