Determining the cause of failure in multi-agent systems is a critical yet unresolved challenge. Existing methods rely on pattern recognition through long conversation logs, resulting in very low step-by-step accuracy. This is due to a lack of robust counterfactual inference—the ability to determine whether a single behavioral modification would have prevented the failure. To bridge this gap in counterfactual inference, this paper presents Abduct-Act-Predict (A2P) Scaffolding, a novel agent framework that transforms failure attribution from pattern recognition to a structured causal inference task. A2P guides a large language model through a formal inference process that involves three steps: 1. Hypothesis formulation: inferring the root cause behind agent behavior; 2. Action: defining a minimal corrective intervention; and 3. Prediction: simulating subsequent trajectories and verifying whether the intervention resolves the failure. Experimental results using the Who-When benchmark demonstrate that A2P achieves step-by-step accuracy that is more than twice that of existing methods.
