ASCon localizes multi-agent failures as one model instead of three, +14.73 Macro-F1 on failure-mode detection.

When a multi-agent run fails you need to know which agent, which step, what kind. ASCon builds contextualized representations from trajectories using direction-aware graph attention, masked step-to-agent attention, and agent-conditioned step contextualization, then attaches specialized heads rather than training separate models (arXiv 2608.10646). +5.83% micro-accuracy on faulty-agent detection, +10.63% on faulty-step, +14.73 Macro-F1 on failure-mode, with better out-of-domain performance when paired with LLM attribution. If you run orchestration in production without a way to localize blame in a trace, this is the shape of what you're missing.