arXiv:2606.20487: H-RePlan — Hierarchical Recovery of AI Agents Across Multiple Devices

Cross-Device AI Agents and the Failure Recovery Problem

Cross-device AI agents — systems that simultaneously manage Linux and Android within a single workflow — are becoming increasingly common in autonomous agent research. Their main weakness is fragility in the face of errors: when one device fails to execute a command, the global orchestrator must replan the entire task from scratch, consuming large numbers of tokens and slowing down execution.

H-RePlan: A Hierarchical Recovery Architecture

H-RePlan, a paper by ten co-authors led by Shu Yao, Yuhu Luo, and Qian Long (published June 19, 2026), addresses this problem with a “compact cross-layer failure abstraction.” Each device has its own local recovery strategy module that resolves minor errors autonomously, without reporting to the orchestrator. Only serious errors that exceed the local recovery capability escalate to the global replanning level.

Concrete Results vs. Baselines

On the new HeraBench benchmark — which injects realistic failures into cross-device workflows — H-RePlan records a higher task completion rate, better adherence to original instructions, and lower token costs compared to global replanning where every error passes through the central orchestrator. HeraBench is the first benchmark specifically designed to test the robustness of cross-device agents under error conditions.

Authors’ Conclusion

The key insight is that “scope-aware” hierarchical recovery is not an optional optimization, but a prerequisite for reliable cross-device AI agents in real-world scenarios.