abstract
The advent of large language models (LLMs) has triggered the emergence of a new class of intelligent systems: generative agents. While promising unprecedented autonomy and flexibility, these agents introduce a fundamental engineering challenge: how to design, control, and make reliable systems whose decision-making core is an inherently stochastic and opaque component. This project proposes a critical and engineering analysis of emerging generative agent architectures, using the established principles of Agent-Oriented Software Engineering (AOSE) as an analytical lens. By examining four representative frameworks - AutoGen, CAMEL, Eclipse LMOS Arc and AgentLite - we aim to demonstrate how each adopts a different engineering strategy to impose structure and predictability on LLM behaviour. The analysis aims to reveal a functional convergence towards AOSE concepts, such as the reinvention of coordination artefacts similar to those of the CArtAgO model. However, it also aims to highlight a significant paradigmatic divergence' from formal organisational models such as Moise, analysing how social structures in generative systems are realised through alternative, often implicit and emergent approaches. It is concluded that a critical understanding of current architectures, informed by AOSE, is a prerequisite for the future synthesis of hybrid architectures that combine the structured deliberation of classical paradigms with the flexibility of LLMs.
outcomes
