XAgent reaches 62.0% on SWE-bench-lite by localizing from dynamic execution behavior instead of the issue text

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.09769 (2026-09-09) argues that agents relying on the static issue description bias their reasoning toward the narrow scope of that text, producing incorrect localization and incomplete validation. XAgent adds dynamic behavioral analysis and extra program context to both localize and validate, reaching a 62.0% resolve rate and 72.8% function-localization accuracy on SWE-bench-lite while staying cost-competitive, and resolving 7 issues that the top existing baselines miss entirely. The direction — execution-guided rather than description-oriented patch generation — is the same instinct behind ExecCritic and AttnCompress this week: the runtime, not the prompt, is where the missing context lives.