A joint software-engineering team from the University of Oulu, the University of L'Aquila, Wageningen University and the University of Southern Denmark filtered 234 industry talks and production cases down to 138 industrial AI agent deployment reports for cross-validated analysis. The study argues the large language model is only the engine, and that whether an enterprise agent ships depends less on prompts than on throttling, retry and fallback, context slicing, state synchronization, tiered memory and centralized permissions and auditing. It distills ten recurring architecture patterns, including the iterative reason-act-observe loop, directed acyclic graph workflows, agents packaged as stateless microservices and hierarchical coordinator designs, and notes that digital SaaS, embodied robotics and cyber-defense settings require fundamentally different architectures. It also warns that each added agent node multiplies communication distortion and cost, and that irreversible actions such as payments or data deletion need physical human-in-the-loop confirmation.

University of Oulu·University of L'Aquila·Wageningen University·+3