The Cognitive Hourglass: Agent Abstractions in the Large Models Era

Recent advances in AI are driving an unprecedented and fastpaced development of myriads of powerful agent tools and applications, mostly based on generative AI technologies such as Large Language/Multi-modal/Agent Models. However, despite many proposals in that direction, the lack of a sound set of usable engineering abstractions hinders the possibility of methodically engineering complex agent-based applications, also due to the gap between cognitive agent-based concepts and LLMs' behavioural patterns. We argue that such a set of abstractions should constitute the narrow neck of an indispensable "cognitive hourglass": a level of abstraction that is meant to be useful for humans to understand/design/control agents and MAS, regardless of the specific AI technologies adopted at the implementation level and of the specific application context. Here, we elaborate on the idea of the cognitive hourglass, motivate its need, sketch its envisioned architecture, and identify the research challenges for its realisation.