A new approach to multi-agent workflow optimization, InFlowOp, is proposed. It uses a label-free cost function to determine task decomposition and agent assignment, and corrects faults during execution with the cheapest move. InFlowOp outperforms single-agent baselines in various domains and backbones.