Designing Domain-Specific Agents via Hierarchical Task Abstraction Mechanism

TL;DR

This work introduces the Hierarchical Task Abstraction Mechanism HTAM for designing domain specific agents by aligning architecture with a domain's intrinsic task dependency graph. HTAM yields a modular, multi layer structure where sub agents operate within a layered pipeline that progresses from data acquisition to synthesis, enforcing procedural correctness. The EarthAgent instantiation demonstrates how HTAM can outperform traditional ReAct, Plan&Execute, Debate and AFlow style systems on GeoPlan bench, a new multi step remote sensing planning benchmark with metrics for tool selection, path similarity, and completeness. Empirical results show robust performance gains across diverse LLM backbones and domains, supporting the claim that task guided hierarchy is a key factor for robust specialized autonomous systems. The work also provides a comprehensive implementation blueprint including task construction, evaluation protocols, and an expansive tool pool to facilitate future HTAM deployments in other domains.

Abstract

LLM-driven agents, particularly those using general frameworks like ReAct or human-inspired role-playing, often struggle in specialized domains that necessitate rigorously structured workflows. Fields such as remote sensing, requiring specialized tools (e.g., correction, spectral indices calculation), and multi-step procedures (e.g., numerous intermediate products and optional steps), significantly challenge generalized approaches. To address this gap, we introduce a novel agent design framework centered on a Hierarchical Task Abstraction Mechanism (HTAM). Specifically, HTAM moves beyond emulating social roles, instead structuring multi-agent systems into a logical hierarchy that mirrors the intrinsic task-dependency graph of a given domain. This task-centric architecture thus enforces procedural correctness and decomposes complex problems into sequential layers, where each layer's sub-agents operate on the outputs of the preceding layers. We instantiate this framework as EarthAgent, a multi-agent system tailored for complex geospatial analysis. To evaluate such complex planning capabilities, we build GeoPlan-bench, a comprehensive benchmark of realistic, multi-step geospatial planning tasks. It is accompanied by a suite of carefully designed metrics to evaluate tool selection, path similarity, and logical completeness. Experiments show that EarthAgent substantially outperforms a range of established single- and multi-agent systems. Our work demonstrates that aligning agent architecture with a domain's intrinsic task structure is a critical step toward building robust and reliable specialized autonomous systems.

Figures (16)

• Figure 1: Comparative analysis of agent architectures on a complex geospatial query. (1) ReAct's iterative, step-by-step process lacks a global strategy, leading to chaotic, redundant tool calls and incomplete solutions. (2) Plan&Execute adheres to a rigid, pre-determined plan, showing no capacity for correction even when encountering errors, ultimately failing the task. (3) HTAM demonstrates a structured, hierarchical decomposition, ensuring a logical progression from data acquisition (Layer 1) to analysis (Layer 2) and final synthesis (Layer 3), leading to a coherent and complete solution. Note that in (1) and (3), each tool-call denotes a thought-action-observation process.
• Figure 2: Conceptual visualization of the planning processes of ReAct, Plan&Execute, and HTAM, inspired by gradient descent optimization. Starting from the same initial state, each architecture navigates the problem space differently. (a) ReAct takes noisy, step-by-step actions, analogous to stochastic gradient descent bottou2012stochastic; its path is highly reactive and can be inefficient. (b) Plan&Execute commits to a single, direct trajectory based on an initial global plan, similar to vanilla gradient descent bishop2006pattern, which can be brittle and fail if the initial direction is flawed. (c) HTAM operates in structured stages, providing a stable yet adaptive path analogous to mini-batch gradient descent bengio2017deep, enabling efficient and robust convergence to the solution.
• Figure 3: Workflow of our EarthAgent. The process starts with a Top-Down Planning, where a user query is decomposed into a hierarchical plan across layers. This is followed by a Bottom-Up Execution, where specialized sub-agents process information.
• Figure 4: The process of task construction and validation. All tasks were built, designed and filtered through this highly-automatic but human-intervened pipeline
• Figure 5: Performance stability of EarthAgent versus ReAct across multiple LLM backbones.