Fractured-vuggy reservoirs exhibit complex spatial structures and significant heterogeneity due to multi-stage tectonic activities. The intricate fracture-vuggy configurations result in unclear well-to-well connectivity channels, making it challenging to determine the injection-production paths. Understanding these connectivity paths is essential for formulating effective injection-production plans and enhancing oil recovery. Traditional research methods often overlook the dynamic response relationship between injection and production wells, leading to low accuracy in describing the spatial distribution characteristics of inter-well connectivity channels. In this study, the multi-media multiphase flow of carbonate rocks with different flow directions is regarded as a multi-agent. A multi-agent deep reinforcement learning model integrating local attention and global constraints is proposed. Based on the dynamic production data, multi-fractal technology is used to describe the dynamic response characteristics between injection and production wells. The connectivity between agents is quantified as the agent action constraint, and a global search strategy is designed. The attention mechanism is introduced to capture the local state information of adjacent agents, deter-mine the optimal search direction of the agent, and automatically search the connected channel of the injection-production well group. Applying this model to a typical unit of the Tahe oilfield, we compare the channel parameters automatically extracted by the inter-well connectivity search with the characteristic parameters of the tracer concentration output curve. Experimental results indicate that the identified paths more accurately describe the spatial distribution characteristics of the connected channels between injection and production wells. This approach provides technical support for adjusting operational systems and improving recovery rates during the water injection development stage of fractured-vuggy reservoirs.