This paper treats failures in long-horizon language-agent rollouts as error-amplification problems in executable planning graphs, rather than isolated bad nodes. It introduces WM-SAR, which estimates node-edge amplification through residual spectral radius, greedily selects a connected repair subgraph using marginal spectral relief, and sends only that region to an LLM for root-cause repair. The abstract reports evaluations on synthetic calling-tree graphs, benchmark-inspired agent topologies, and cross-model LLM repair experiments, where WM-SAR improves long-horizon stabilization and root-cause recovery under compact token budgets.
No heat snapshots are available in the last 24 hours.