Back to AI Research

AI Research

Atomic Task Graph: A Unified Framework for Agentic... | AI Research

Key Takeaways

  • Atomic Task Graph: A Unified Framework for Agentic Planning and Execution introduces a new way for AI agents to handle complex, multi-step tasks.
  • LLM-based agents have shown strong potential for solving complex multi-step tasks, yet existing performance improvements often rely on either scaling to larger backbone models or task-specific fine-tuning.
  • The former incurs substantial computational costs, while the latter typically generalizes poorly across different tasks.
  • To address these limitations, we propose Atomic Task Graph (ATG), a unified control framework for planning and execution.
  • Specifically, ATG maintains an explicit graph to expose dependencies and support reuse.
Paper AbstractExpand

LLM-based agents have shown strong potential for solving complex multi-step tasks, yet existing performance improvements often rely on either scaling to larger backbone models or task-specific fine-tuning. The former incurs substantial computational costs, while the latter typically generalizes poorly across different tasks. Although prompt-based control is training-free and broadly applicable, existing methods still leave input-output dependencies between subtasks implicit in textual trajectories, making verified intermediate results difficult to reuse. To address these limitations, we propose Atomic Task Graph (ATG), a unified control framework for planning and execution. Specifically, ATG maintains an explicit graph to expose dependencies and support reuse. During planning, it recursively decomposes a high-level task into subtasks, forming a sequence of directed acyclic graphs (DAGs) whose evolution can be traced. During execution, the dependencies exposed by ATG allow independent branches to be executed in parallel, thereby improving execution efficiency. When failures are detected, ATG leverages the graph evolution history to localize the error source and repair only the affected region, preserving validated regions unchanged. Experiments show that ATG consistently outperforms strong baselines in success rate and execution efficiency across three interactive benchmarks using only 7B-8B backbones.

Atomic Task Graph: A Unified Framework for Agentic Planning and Execution introduces a new way for AI agents to handle complex, multi-step tasks. Current AI agents often struggle with long-term planning because they rely on linear, text-based sequences, which makes it difficult to track dependencies between steps or recover from errors without starting over. This framework solves these issues by organizing tasks into a structured graph, allowing the agent to plan, execute, and repair its work more efficiently.

From Linear Text to Structured Graphs

Most AI agents solve problems by generating a long, linear chain of text. This approach is prone to errors because the agent can lose track of how different steps depend on one another, and the growing amount of text can lead to "hallucinations" or incorrect actions. The Atomic Task Graph (ATG) framework replaces this linear process with a Directed Acyclic Graph (DAG). In this structure, every node represents an atomic tool-use unit, and the edges between them explicitly define the input-output dependencies. This allows the system to see exactly which parts of a task rely on others.

How the Framework Works

The ATG process consists of three main stages:

  • Recursive Graph Compilation: The agent breaks down a high-level goal into smaller, manageable subtasks. It does this recursively, ensuring that each refined step preserves the original input-output interface. This creates a clear, traceable history of how the plan was formed.

  • Dependency-Aware Execution: Because the graph explicitly shows dependencies, the agent can identify which tasks are independent and execute them in parallel, significantly speeding up performance. Before running the plan, the agent performs a "thought experiment" to simulate the process and catch potential failures early.

  • Minimal Necessary Subgraph Repair: If an error occurs during execution, the agent does not need to restart the entire task. Instead, it uses the graph’s history to pinpoint the exact node where the failure originated and repairs only that specific region, keeping the rest of the validated work intact.

Performance and Efficiency

Experiments across three benchmarks—ALFWorld, WebShop, and ScienceWorld—show that ATG significantly outperforms traditional methods. By using 7B–8B parameter models, the framework achieves higher success rates than existing approaches that rely on larger, more expensive models. The explicit graph structure also reduces the number of steps required to complete a task and lowers the rate of hallucinated actions, as the agent no longer needs to reason over an ever-growing, unstructured text history.

Key Takeaways

The success of ATG demonstrates that the way an agent organizes its work is just as important as the size of the underlying model. By treating the task as a graph rather than a text sequence, the framework provides a reliable substrate for validation and recovery. The ability to localize repairs and execute independent branches in parallel makes the system both more robust and faster, proving that structured control mechanisms can bridge the performance gap between smaller open-source models and larger, proprietary ones.

Comments (0)

No comments yet

Be the first to share your thoughts!