Towards an Approach for Specifying Intelligent Systems Involving Foundation Model Based Agents

The rise of foundation-model-based agents has introduced a new generation of intelligent systems capable of autonomous reasoning, reflection, and coordination. However, their specification remains challenging due to emergent behaviors and dynamic, probabilistic interactions that exceed the assumptions of traditional software and requirements engineering approaches. This short paper presents the vision of PerSpecAgents, a perspective-based approach designed to support the specification of intelligent systems involving such agents. The solution proposal extends an ML-enabled system specification approach by incorporating agent-oriented concerns, such as autonomy, reasoning loops, reflection mechanisms, memory management, and ethical guardrails. An initial academic validation with graduate-level participants allowed refinements and provided a preliminary indication of PerSpecAgents' clarity, completeness, and usefulness. These findings strengthen our confidence in the approach's potential to support the specification of intelligent systems built upon foundation models.