CAST turns sparse task outcomes into action-level critique training for long-horizon tool-calling agents

arXiv 2608.30147·medium signalXBlueskyLinkedInCopy link

arXiv 2608.30147 (submitted 2026-08-31) starts from the case where a single wrong action, like refunding the wrong purchase, causes irreversible failure and has to be intercepted before execution. Rather than the prompt-based critique agents most systems use, CAST analyzes agent trajectories to synthesize structured rationales explaining why an action is valid or invalid under partial observability and domain policy, then uses those for both critique learning and policy optimization. The framing that reliability must hold across repeated trials, not just per run, is the part worth stealing for anyone running agents against production side effects.