Atomic Task Graph Boosts LLM Agent Planning and Execution Efficiency

Large Language Model (LLM)-based agents show promise for complex multi-step tasks, but performance gains often come from scaling to larger, more expensive models or task-specific fine-tuning that lacks generalization. Existing prompt-based control methods also struggle with implicit input-output dependencies, making intermediate result reuse difficult. To overcome these limitations, researchers introduce the Atomic Task Graph (ATG), a unified control framework for agentic planning and execution. ATG explicitly maintains a graph structure to reveal dependencies between subtasks, enabling better reuse of verified intermediate results. During planning, ATG recursively decomposes high-level tasks into subtasks, forming traceable directed acyclic graphs (DAGs). For execution, these explicit dependencies allow independent branches to run in parallel, significantly boosting efficiency. Furthermore, ATG leverages its graph history to localize and repair errors only in affected regions, preserving validated parts. Experiments demonstrate that ATG consistently outperforms strong baselines in success rate and efficiency across interactive benchmarks, even when using smaller 7B-8B LLM backbones.