Reliable Intention Selection in BDI Agents with Recovery Shields

Safety shields have emerged as a practical way to enforce runtime properties and handle failure in autonomous agents. The existing approaches mainly focus on single-agent systems and respond to violations by immediately rejecting unsafe actions. In this paper, we extend the notion of safety shields for Belief-Desire-Intention agents by proposing a revised version where a shield can observe and respond not only to what the agent itself does, but also to actions and changes from other agents, when these affect the shielded agent’s behaviour. Our shields suspend intentions that would break a formal specification and resume them once it is safe to do so. To support this, we introduce recovery shields, a new mechanism that defines when a suspended intention can be safely resumed. By bringing both enforcement and recovery into the agent reasoning cycle, our framework improves the reliability of multi-agent systems while preserving autonomy and flexibility. Our main contributions are extensions of the reasoning cycle and operational semantics of AgentSpeak(L) and an implementation in the JaCaMo platform.