Let Me Talk to You! Natural Language Interaction Between Humans and BDI Agents via ChatBDI

In this paper we describe ChatBDI, a framework for extending Belief-Desire-Intention (BDI) agents implemented in Jason with the ability to understand and generate messages in natural language by exploiting embeddings and Large Language Models (LLMs). Thanks to the generative power of LLMs, the ‘chattification’ of new or legacy multiagent systems (MAS) adds a creative and fluent ‘language actuator’ to BDI agents and serves two main purposes. First, it allows users to enter the MAS and conversate with any other software agent in natural language, as if humans were agents themselves. Second, it allows both the MAS developers and the users to follow the conversation among software agents and to ask them information on their behavior and decisions, acting as a lightweight co-pilot and improving transparency and explainability. The major strength of ChatBDI, and its distinguishing feature w.r.t. related works, is its general purpose nature: by exploiting plan injection and sophisticated meta-programming facilities, ChatBDI can be used to chattify any MAS implemented in Jason or JaCaMo with neither adaptations of the ChatBDI code itself, nor changes to the existing AgentSpeak(L) agents’ source code.