Industrial Internet of Things (IIoT) is a key technological pillar of the Fourth Industrial Revolution, also known as Industry 4.0. In this context, an area of considerable interest is human safety technology. Solutions rely on multiple sensors connected to a central monitoring system and supported with software to autonomously or semi-autonomously identify safety hazards. To this end, Computer vision systems are leveraged. However, streaming continuous video from numerous sensors can strain network resources, risking timely hazard response in large industrial setups. This work proposes a reference IIoT architecture based on Multi-Agent Systems to manage safety risks. It allows for scalable sensor integration and dynamically assesses sensor input based on risk levels. To prevent network overload, the architecture uses sensor-level intelligence at the edge layer to assess situational risks and decide whether to forward video signals to a centralized local cloud agent. The central cloud agent, using strategies like ensemble learning, selectively requests additional data from distributed edge agents based on the diagnosed risk. This approach was tested in monitoring safety during aircraft assembly, showing that edge processing reduces network load by limiting unnecessary data transmission without compromising accuracy. This architecture effectively distributes processing to the edge, maintaining detection accuracy while minimizing network traffic compared to continuous centralized video transmission.