Bridging Symbolic and Generative AI for Engineering Agents: An AOSE-Driven Evaluation of BDI and LLM-based Paradigms

The evolution of software architectures for agent systems is characterised by the transition from deterministic symbolic paradigms, such as the Belief-Desire-Intention (BDI) model, to probabilistic approaches based on Large Language Models (LLMs). However, the integration of these models into control cycles introduces engineering challenges related to deliberative opacity, hallucinations, and the difficulty of ensuring intention persistence in dynamic environments. This thesis addresses the problem of reconciling the flexible autonomy of generative agents with the structural rigour required for the development of robust software systems. The research adopts a methodology divided into distinct phases: a critical analysis of the main generative frameworks (AutoGen, CAMEL, Eclipse LMOS, AgentLite); the transposition of Russell & Norvig's basic agent program designs into AutoGen, with the formalisation of the PerceptDriven (structured) and LLMDriven (purely generative) architectural patterns; and the comparative design of advanced architectures, contrasting a symbolic agent developed in ASTRA language with a neuro-symbolic agent inspired by the CoALA framework, equipped with structured long-term memory. The experimental verification, conducted in the Dynamic Vacuum World stochastic scenario, highlights an engineering trade-off between controllability and flexibility. The results show that, while the purely generative approach has limitations in memory management and long-term goal setting, the neuro-symbolic architecture offers superior cognitive resilience and better error management, albeit introducing computational latencies compared to the stability of symbolic systems. The work concludes by proposing a convergence towards hybrid architectures, in which the symbolic core manages procedural reactivity and the LLM operates as Cognitive Failover for planning in unexpected contexts.