arXiv cs.AIMonday · May 25, 2026FREE

DART: Semantic Recoverability for Structured Tool Agents

The arXiv paper "DART: Semantic Recoverability for Structured Tool Agents" introduces a novel runtime system aimed at resolving the complex issue of mid-execution failures in structured tool agents. When such an agent fails, current recovery methods present a dilemma: replaying the entire task is safe but inefficient, while restoring from a local checkpoint risks leaving committed downstream work tied to an outdated upstream history. This problem is particularly acute in "commitment-sensitive" scenarios where downstream consumers have already acted on an agent's output, and a simple rollback of the failed instance could invalidate subsequent operations. DART formalizes this challenge as "semantic recoverability." The system operates by localizing the failed instance, certifying semantically recoverable boundaries within that instance, and aligning checkpoints to these boundaries. It then intelligently selects an "admissible restore point" that ensures committed downstream work remains valid, respecting dependency and effect constraints. If a valid restore point cannot be found, DART blocks further execution to prevent data inconsistency. The paper reports that DART was validated across three distinct LLM-driven domains and externally on a LangGraph-based substrate, demonstrating its ability to correctly recover all evaluated commitment-sensitive cases where baseline local recovery approaches failed. This research, published on arXiv on May 25, 2026, offers a significant step towards more robust and reliable AI agent architectures.