Active digital agents (Idas) provides for Apis officers to perform functions by responding to the user requests. While Idas is enabled in large teaching models (llms) to respond to the reply from interface reduction in multi-step exchanges, trained in their digital locations. Previous ways are achieving under part of the activities in complex benchmarks such as AppWorld. We present a learning method (RL) that train Idas directly to their target areas. We officially set this training as a visual markov …
