Explainable Agents (XAg) by Design

Sebastian Rodriguez, John Thangarajah

Abstract: The likes of ChatGPT has propelled the use of AI techniques beyond our community's expectations. Along with this, the fear of AI has also risen, in particular around the ability, or lack thereof, of the AI system to explain its behaviours. Explainability is a key element of building trust and an important issue for our community. In this paper we advocate for agents that are explainable-by-design, that is, explainability is built into the development of agents rather than an afterthought. We propose key features of an explainable agent (XAg) system and propose a general framework that enables explainability. We advocate the use of design patterns to develop XAgs and propose a general design pattern that can be used for any agent architecture. We instantiate our framework for goal-based agents and implement the framework for the SARL agent programming language coupled with a state-of-the-art event management system. We make a call to the developers of other agent programming languages (APLs) in our community to follow suit by instantiating the general framework we propose into their APL, perhaps even enhancing the framework we present. We also propose an open repository of design patterns and examples for agent systems. If nothing else, we hope this paper will inspire further work on XAg from the design perspective as it is critical that multi agent systems are explainable by design!