Beyond Global Replanning: Hierarchical Recovery for Cross-Device Agent Systems

2026-06-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

H-RePlan is a novel hierarchical replanning framework designed for multi-device agent systems, addressing the limitations of coarse-grained recovery in existing solutions. It equips each device with interchangeable API, CLI, and GUI execution strategies, separating device-local strategy recovery from orchestrator-level global replanning via a compact Cross-Layer Failure Event (CLFE) abstraction. Evaluated on HeraBench, a fault-injected benchmark comprising 23 seed tasks expanded into 174 variants across Linux and Android devices, H-RePlan significantly outperforms baselines like UFO3-GUI. It achieves 75.84% completion, 77.72% instruction adherence, and a 36.78% perfect-pass rate, drastically reducing the expected token cost per perfect pass to 1.93M tokens from UFO3-GUI's 10.51M tokens.

Key takeaway

For AI Engineers developing multi-device automation, H-RePlan demonstrates that robust agent systems require a hierarchical recovery architecture. You should consider implementing distinct device-local strategy planners and a global orchestrator, using a structured failure abstraction like CLFE to guide efficient recovery. This approach significantly improves task completion, adherence, and cost-efficiency, especially under dynamic runtime failures across heterogeneous environments.

Key insights

Hierarchical recovery, separating local strategy revision from global replanning, is crucial for robust multi-device agent systems.

Principles

• Separate device-local strategy recovery.
• Orchestrator handles cross-device replanning.
• Unified API–CLI–GUI execution is vital.

Method

H-RePlan uses an Orchestrator for global task planning and device-level Strategy Planners for local execution and recovery, communicating via Cross-Layer Failure Events (CLFE) to manage API, CLI, and GUI strategies.

In practice

• Implement local strategy switching for failures.
• Design structured failure abstraction (CLFE).
• Integrate API, CLI, and GUI agents.

Topics

• Multi-device Agents
• Hierarchical Planning
• LLM Agents
• Fault Tolerance
• API-CLI-GUI Control
• HeraBench

Code references

• openclaw/openclaw

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Beyond Global Replanning: Hierarchical Recovery for Cross-Device Agent Systems — Hierarchical, scope-aware recovery is crucial for robust multi-device agent systems operating under dynamic runtime failures.
• ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies — ReCoVLA uses a VLM to compile structured rewards for residual policy training, enabling robust VLA policy recovery.
• Self-Healing Agentic Orchestrators for Reliable Tool-Augmented Large Language Model Systems — Self-healing orchestrators improve LLM system reliability by budgeting recovery actions and verifying outcomes.
• When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents — LLM agents struggle with dynamic replanning and error recovery when tools fail, a problem unaddressed by model scaling.
• Article: Building Hierarchical Agentic RAG Systems: Multi-Modal Reasoning with Autonomous Error Recovery — Hierarchical multi-agent RAG with autonomous error recovery significantly improves accuracy and reduces hallucinations across structured and unstructured enterprise data.
• CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation — An agent harness with model-driven recovery significantly improves reliability and autonomy in LLM-driven engineering simulations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.