RS-Claw: Progressive Active Tool Exploration via Hierarchical Skill Trees for Remote Sensing Agents

# RS-Claw: Progressive Active Tool Exploration via Hierarchical Skill Trees for Remote Sensing Agents
Source: [https://arxiv.org/html/2605.13391](https://arxiv.org/html/2605.13391)
Liangtian Liu, Zeyuan Wang, Ziyu Li, Kai Ouyang, Zichao Tang, Chengfu Liu, Haifeng Li, Hanwen Yu, Wentao Yang, Cheng Yang, and Dongyang Hou\(Corresponding authors: Cheng Yang and Dongyang Hou\.\)Liangtian Liu, Zeyuan Wang, Ziyu Li, Kai Ouyang, Zichao Tang, Chengfu Liu, Haifeng Li, Cheng Yang, and Dongyang Hou are with the School of Geosciences and Info\-Physics, Central South University, Changsha 410083, China \(e\-mails: llt62t@csu\.edu\.cn; 255011039@csu\.edu\.cn; liziyucsu@csu\.edu\.cn; oykkksk@csu\.edu\.cn; 2121050207@csu\.edu\.cn; 8211220119@csu\.edu\.cn; lihaifeng@csu\.edu\.cn; ychades@csu\.edu\.cn; houdongyang1986@csu\.edu\.cn\)\.Hanwen Yu is with the School of Resources and Environment, University of Electronic Science and Technology of China, Xian 710071, China \(e\-mail: yuhanwenxd@gmail\.com\)\.Wentao Yang is with the School of Earth Sciences and Spatial Information Engineering, Hunan University of Science and Technology, Xiangtan 411201, China, and also with the Sanya Institute of Hunan University of Science and Technology, Sanya 572024, China \(e\-mail: yangwentao8868@126\.com\)\.

###### Abstract

The rise of multi\-modal large language model \(MLLM\) led remote sensing \(RS\) intelligence to a new paradigm shift, i\.e\., from “see” to “action”, especially, OpenClaw\-style frameworks promise charming ability that autonomously operate massive RS image processing tools to execute complex tasks\. Existing RS agents adopt a passive selection paradigm for tool invocation, relying on either full tool registration \(Flat\) or retrieval\-augmented generation \(RAG\)\. However, when confronted with the massive and multi\-source heterogeneous RS tool ecosystem, such passive mechanisms struggle to dynamically balance “context load” and “toolset completeness” throughout the task reasoning process, thus exhibiting inherent limitations: full tool registration triggers context space deficits during long\-horizon tasks, whereas RAG retrieval leads to the omission of critical tools in essential steps\. To overcome these bottlenecks, this paper redefines the tool selection paradigm—arguing that the agent should act as an active explorer within the tool space\. Based on this perspective, we propose RS\-Claw, a novel RS agent architecture\. By leveraging Skill encapsulation technology at the tool end, this architecture hierarchically structures tool descriptions, enabling the agent to execute on\-demand sequential decision\-making: initially selecting relevant skill branches by reading only tool summaries, then dynamically loading and reading detailed descriptions, and ultimately achieving precise invocation\. This active paradigm not only significantly liberates the agent’s context space but also effectively ensures the accurate hit rate of critical tools during long\-horizon reasoning\. Systematic experiments on the Earth\-Bench benchmark demonstrate that RS\-Claw’s active exploration mechanism effectively filters semantic noise and substantially frees up reasoning space \(achieving an input token compression ratio of up to 86%\), comprehensively outperforming existing Flat and RAG baselines across various complex reasoning evaluations\.

## IIntroduction

In recent years, driven by the continuous advancements of large language models \(LLMs\) in intent comprehension, autonomous decision\-making, and tool utilization, researchers have developed LLM\-based agents equipped with specialized toolsets to autonomously execute complex reasoning tasks\. These agents are capable of parsing natural language instructions and executing tasks through autonomous planning and the invocation of external tools\[[39](https://arxiv.org/html/2605.13391#bib.bib2),[31](https://arxiv.org/html/2605.13391#bib.bib3),[33](https://arxiv.org/html/2605.13391#bib.bib4),[40](https://arxiv.org/html/2605.13391#bib.bib28)\]\. In the general domain, for instance, agents such as OpenClaw\[[27](https://arxiv.org/html/2605.13391#bib.bib38)\]have demonstrated significant potential in accomplishing tasks via dynamic tool invocation\. Capitalizing on the autonomous task\-execution capabilities of these agents, researchers in the remote sensing \(RS\) community have leveraged them to process acquired imagery according to diverse customized requirements\. For example, RS\-Agent\[[42](https://arxiv.org/html/2605.13391#bib.bib5)\]achieves the automated execution of RS tasks across various scenarios—including scene classification, visual question answering \(VQA\), and object counting—through task decomposition and the collaborative scheduling of multiple tools\. Furthermore, Earth\-Agent\[[7](https://arxiv.org/html/2605.13391#bib.bib6)\]integrates domain\-specific tools for index inversion, target recognition, and statistical analysis, thereby facilitating multimodal quantitative spatiotemporal reasoning and scientific analysis\.

Most existing RS agents adopt the “full tool registration \(Flat\)” paradigm prevalent in the general domain, wherein the functional descriptions of all candidate tools are directly concatenated into the system prompt\. However, the complex workflows of RS data processing and the rapid iteration of multi\-source heterogeneous data\[[5](https://arxiv.org/html/2605.13391#bib.bib1)\]necessitate that these agents be equipped with massive and dynamically expanding tool libraries containing hundreds of functions, such as QGIS, GDAL, Google Earth Engine, and Orfeo ToolBox\[[8](https://arxiv.org/html/2605.13391#bib.bib39),[9](https://arxiv.org/html/2605.13391#bib.bib40)\]\. Confronted with such ultra\-large\-scale toolsets, the traditional Flat paradigm inevitably induces a context space bottleneck within LLMs\[[29](https://arxiv.org/html/2605.13391#bib.bib7)\], precipitating two critical challenges during the agent’s task reasoning:

1. 1\.Constrained reasoning space in long\-horizon tasks\. Taking the Earth\-Bench evaluation benchmark as an example, the full registration of its 104 domain\-specific RS tools and their parameter specifications alone consumes upwards of 20k tokens\. Although several state\-of\-the\-art LLMs support extended contexts, lengthy tool descriptions inevitably occupy a substantial portion of the context space\. This significantly compresses the essential context space required by the agent for storing intermediate state data, conducting multi\-step trial\-and\-error, and executing ReAct reasoning during long\-horizon tasks\[[14](https://arxiv.org/html/2605.13391#bib.bib27),[2](https://arxiv.org/html/2605.13391#bib.bib41),[43](https://arxiv.org/html/2605.13391#bib.bib13)\]\.
2. 2\.Semantic confusion within multi\-source tool spaces\. Given the disparate physical mechanisms and algorithms associated with different sensors, voluminous and semantically similar tool descriptions generate extensive “semantic noise\.” Consequently, the model suffers from “attention defocus” and the “lost in the middle” phenomenon, rendering it highly susceptible to “tool hallucination” when parsing user intents\[[24](https://arxiv.org/html/2605.13391#bib.bib8),[28](https://arxiv.org/html/2605.13391#bib.bib21)\]\.

![Refer to caption](https://arxiv.org/html/2605.13391v1/fig1.png)Figure 1:Comparison of agent tool selection paradigms\. \(a\) Passive paradigm: Existing methods define the agent as a passive tool recipient\. Specifically, the “Flat” strategy globally injects all tools, which leads to context space overflow and severely compresses the reasoning space; the “RAG” strategy, despite freeing up some space, is susceptible to causing the omission of critical tools during long\-horizon reasoning\. \(b\) Active paradigm: RS\-Claw defines the agent as an active tool explorer\. Leveraging a hierarchical skill tree, the agent progressively discovers and loads tools on demand through sequential decision\-making\. This mechanism, which interleaves dynamic tool loading with ReAct reasoning, simultaneously ensures ample context reasoning space and an accurate tool hit rate\.To mitigate the context bottleneck induced by Flat paradigm, existing research has primarily explored tool registration paradigms based on external retrieval augmentation \(RAG\)\[[16](https://arxiv.org/html/2605.13391#bib.bib9)\]\. By filtering a subset of candidate tools from a large\-scale tool library prior to task execution, such methods reduce the context space occupied by tool descriptions to a certain extent \(e\.g\., ToolReAgt\[[3](https://arxiv.org/html/2605.13391#bib.bib10)\]\)\. However, this approach still exhibits significant limitations in RS scenarios\. RS tasks frequently involve multi\-stage task decomposition and state\-dependent tool requirements\. Relying solely on the initial task description and surface\-level semantic similarity for a single\-shot retrieval makes it difficult to adequately cover the critical tools required in subsequent reasoning stages\. Consequently, although the RAG paradigm alleviates the context load issue, it may sacrifice the completeness of the toolset, making it challenging for the agent to simultaneously balance context efficiency and multi\-step reasoning completeness within a large\-scale tool space\.

Consequently, existing methods still struggle to simultaneously balance exploration efficiency, reasoning completeness, and system scalability within large\-scale tool spaces\. To break through these bottlenecks, we argue that an agent’s tool invocation should mirror human cognitive logic: tool acquisition should neither be a globally static passive retrieval nor a blind code generation process\. Instead, it should be a task\-driven, active exploration and on\-demand discovery process grounded in contextual reasoning\. To actualize this mechanism, this paper fundamentally reconstructs the tool\-side architecture of the agent, proposing a novel framework named RS\-Claw\. Inspired by the concept of “progressive disclosure,” we discard the traditional flattened tool list\. Instead, leveraging expert knowledge, we semantically encapsulate the voluminous RS toolset into “skills,” thereby constructing a structured, hierarchical skill tree\[[1](https://arxiv.org/html/2605.13391#bib.bib12)\]\. Under this architecture, we innovatively model the discovery and selection of tools as an autonomous sequential decision\-making process executed by the agent\. Rather than relying on externally predefined program logic for local registration, the agent incorporates “tool exploration” as an intrinsic action within its decision space\. It is thus guided to conduct progressive, autonomous searches and dynamic loading within the hierarchical tree\. This “active exploration” paradigm effectively resolves the two aforementioned challenges: First, the on\-demand loading of tools reduces the context consumption from a globalO​\(N\)O\(N\)to a localO​\(K\)O\(K\), thereby freeing up ample space for the agent to maintain complex intermediate geospatial states and execute multi\-step logical reasoning\. Second, intent\-driven hierarchical exploration filters out extraneous information early in the decision\-making pathway, thereby eliminating “semantic noise” within the multi\-source tool space\. This effectively mitigates the issues of attention defocus and tool hallucination\.

The main contributions of this article are summarized as follows:

1. 1\.We introduce a pioneering perspective on tool selection for RS agents\. Inspired by human cognitive logic, we redefine the tool selection process as an “active exploration,” thereby transforming the agent from a passive tool recipient into an active explorer within a structured tool space\. Through this theoretical shift, we systematically analyze and address the critical challenges of constrained reasoning space and semantic confusion that agents traditionally encounter under the Flat paradigm\.
2. 2\.Inspired by the concept of progressive disclosure, we design a hierarchical skill tree at the tool level, innovatively internalizing the tool selection action into the agent’s autonomous sequential decision\-making process\. This architecture enables the agent to dynamically load tool subsets on demand, which not only liberates the reasoning context space but also effectively filters out the semantic confusion and information noise inherent in multi\-source toolsets\.
3. 3\.We conduct systematic experiments on the Earth\-Bench benchmark\. The proposed RS\-Claw comprehensively outperforms both the Flat and RAG baselines across all models and evaluation modes, achieving an improvement of 12\.45% over the Flat baseline in the Qwen3\-32b AP mode\. In terms of context space optimization, RS\-Claw significantly reduces the input token count per round compared to the Flat paradigm, achieving an input token compression ratio of up to 86% in the Qwen3\-32b AP mode\. Extensive ablation and scalability experiments further validate the structural rationale and robustness of our proposed method\.

## IIRelated Work

Autonomous agents driven by LLMs are progressively reshaping the automation paradigms for scientific computing and Earth observation tasks\. Existing research focuses on system architectures with language models as the core controller, integrating multi\-step planning, state memory, and tool invocation mechanisms\. These systems have demonstrated robust reasoning and decision\-making capabilities in complex open\-ended tasks, driving the evolution of agents from single\-model inference to multi\-tool collaborative execution\. This paradigm provides a vital foundation for building agents in vertical domains; accordingly, RS tasks have transitioned from multimodal language model adaptation to autonomous systems capable of tool scheduling\. However, as the scale of available tools continues to expand, the effective organization and efficient retrieval of large\-scale tools during task execution have become critical factors influencing reasoning efficiency and task completion quality\. In the context of growing toolsets and increasingly complex task chains, the tool acquisition process still primarily relies on predefined scopes or static single\-step decisions, offering limited support for the dynamic evolution of tool requirements during execution\. This challenge is particularly pronounced in the field of RS, which is characterized by highly specialized tool ecosystems and complex task dependencies\. As RS agents evolve toward increasingly complex long\-horizon tasks, the rapid expansion of available tools has further made efficient tool organization and acquisition a critical challenge\. To address this issue, our work further focuses on dynamic tool discovery in large\-scale tool spaces to support on\-demand tool exploration for complex RS tasks\.

### II\-AFrom General\-Purpose Agent Systems to RS Task Automation

In recent years, the paradigm of building autonomous agents with Large Language Models \(LLMs\) as the core controller has advanced rapidly, becoming a mainstream technical route for complex task automation\. Within this framework, the LLM functions as a unified decision\-making hub, advancing task execution through sub\-goal decomposition, task planning, and the coordination of external tool calls\[[40](https://arxiv.org/html/2605.13391#bib.bib28)\]\. The ReAct\[[43](https://arxiv.org/html/2605.13391#bib.bib13)\]framework introduced the interleaved generation of reasoning traces and action decisions, enabling the model to continuously update plans and retrieve external information in dynamic environments, thereby significantly enhancing problem\-solving capabilities for complex tasks\. Building on this, Reflexion\[[34](https://arxiv.org/html/2605.13391#bib.bib14)\]introduced a linguistic self\-reflection mechanism, allowing agents to formulate improvement strategies by summarizing past failure trajectories, while Tree of Thoughts\[[44](https://arxiv.org/html/2605.13391#bib.bib42)\]studied explicit search over intermediate reasoning paths for complex problem solving\.

Driven by this general paradigm, research on task automation in the field of RS has progressively evolved from single\-model inference toward Vision\-Language Model \(VLM\) systems geared for multi\-task interaction\. Early research primarily focused on adapting VLMs to RS scenarios to achieve foundational capabilities such as image captioning, visual question answering \(VQA\), and spatial understanding\. For instance, RSGPT\[[11](https://arxiv.org/html/2605.13391#bib.bib15)\]constructed high\-quality RS image\-text datasets and fine\-tuned existing VLMs to enable image description and VQA capabilities\. GeoChat\[[13](https://arxiv.org/html/2605.13391#bib.bib16)\]further introduced region\-level inputs and spatial coordinate representation mechanisms, allowing the model to perform region\-level reasoning and visual grounding, thus supporting more granular spatial interactions\. EarthGPT\[[45](https://arxiv.org/html/2605.13391#bib.bib17)\]integrated multi\-source sensor data through a unified instruction tuning strategy, achieving unified modeling of multi\-task RS interpretation\. Building on this, EarthDial\[[36](https://arxiv.org/html/2605.13391#bib.bib18)\]extended processing to multi\-temporal, multi\-spectral, and multi\-resolution RS data, supporting complex temporal analysis and change detection tasks\. RemoteCLIP\[[23](https://arxiv.org/html/2605.13391#bib.bib29)\]established a vision\-language foundation model pre\-trained specifically on RS imagery, providing strong cross\-modal alignment for downstream RS tasks, while SkySense\[[10](https://arxiv.org/html/2605.13391#bib.bib43)\]further explored multi\-modal RS foundation modeling\. However, these systems remain largely instruction\-driven and are primarily oriented toward single\-step or weakly\-planned tasks, showing certain limitations in practical applications that require complex process control and specialized tool invocation\.

To enhance RS task automation, researchers have begun exploring RS autonomous agent systems with LLMs as the core controller\. Early work such as RS\-Agent\[[42](https://arxiv.org/html/2605.13391#bib.bib5)\]initially validated the feasibility of LLM\-driven RS agents; by using an LLM as the central controller combined with RAG and specialized knowledge bases, it achieved automated execution of RS tasks—including scene classification, visual question answering, and object counting—through task decomposition and multi\-tool collaborative scheduling\. Subsequently, Earth\-Agent\[[7](https://arxiv.org/html/2605.13391#bib.bib6)\]integrated 104 domain\-specific tools to enable multimodal quantitative spatiotemporal reasoning and scientific analysis\. For interactive change analysis, ChangeAgent\[[22](https://arxiv.org/html/2605.13391#bib.bib30)\]introduced an RS agent framework that combines temporal\-image interpretation with user interaction and change reasoning\. Addressing task decomposition and cross\-domain coordination in complex workflows, GeoLLM\-Squad\[[15](https://arxiv.org/html/2605.13391#bib.bib19)\]introduced a multi\-agent collaboration strategy, where an orchestrator breaks down user requests into sub\-tasks for specialized domain agents to process, thereby improving modularity and scalability\. Regarding evaluation, Earth\-Agent\[[7](https://arxiv.org/html/2605.13391#bib.bib6)\]introduced the Earth\-Bench benchmark, while GeoLLM\-QA\[[35](https://arxiv.org/html/2605.13391#bib.bib31)\]and ThinkGeo\[[32](https://arxiv.org/html/2605.13391#bib.bib20)\]established benchmark settings for tool\-augmented agents and real\-world RS imagery\. Together, these efforts have propelled a paradigm shift in RS agents from single workflows to multi\-source scientific analysis\.

Although the aforementioned studies have advanced RS tasks from single\-step perception toward the automated execution of complex workflows, most existing RS agents still assume that the tool space is predefined and statically accessible—that is, the agent can access the complete set of tools before task execution begins\. As the number of specialized tools in the RS domain continues to grow and task pipelines become increasingly long and complex, this fixed tool exposure paradigm gradually faces challenges such as high context load and declining tool selection efficiency\. To address this issue, this paper further investigates tool organization and acquisition in large\-scale RS tool environments, and explores a more dynamic tool access mechanism that better aligns with the demands of complex tasks\.

### II\-BAgent Tool Organization and Retrieval

As the number of invokable tools for agents continues to grow, the efficient organization and dynamic acquisition of these tools have become critical factors influencing system performance and scalability\. Early research typically adopted a Flat paradigm, where all tool descriptions were loaded into the context at once before task execution, allowing the language model to select and invoke them directly\. While effective for small\-scale toolsets, this approach faces rapidly escalating context load and attention dispersion issues as the number of tools expands to tens of thousands\[[30](https://arxiv.org/html/2605.13391#bib.bib32),[26](https://arxiv.org/html/2605.13391#bib.bib33)\]\. Consequently, these challenges limit the stability and reasoning efficiency of systems in complex task scenarios\. The scale of this challenge has grown substantially: benchmarks such as API\-Bank\[[19](https://arxiv.org/html/2605.13391#bib.bib34)\], ToolBench\[[29](https://arxiv.org/html/2605.13391#bib.bib7)\], and Tool Decathlon\[[18](https://arxiv.org/html/2605.13391#bib.bib35)\]evaluate whether and how LLMs select and invoke external tools across increasingly large\-scale tool libraries and long\-horizon tasks, while system works such as TaskMatrix\.AI\[[21](https://arxiv.org/html/2605.13391#bib.bib36)\]demonstrate modular tool/API orchestration over heterogeneous components\.

To alleviate the context load imposed by large\-scale tool spaces, researchers have proposed tool selection strategies based on external retrieval mechanisms\. These methods employ a retriever to measure the semantic relevance between tool descriptions and sub\-tasks, filtering a set of candidate tools from the library for the language model’s final decision\. For instance, Gorilla\[[28](https://arxiv.org/html/2605.13391#bib.bib21)\]integrated Retrieval\-Aware Training \(RAT\) into the fine\-tuning process to mitigate hallucinations in tool calling\. ToolLLM\[[29](https://arxiv.org/html/2605.13391#bib.bib7)\]constructed ToolBench, a massive benchmark containing over 16,000 real\-world REST APIs, and introduced a neural API retriever to support the recall of relevant tools within extensive tool spaces, enabling automated selection for complex tasks\. AnyTool\[[6](https://arxiv.org/html/2605.13391#bib.bib22)\]implemented a hierarchical API retrieval structure to improve efficiency by narrowing the search scope layer by layer\. Re\-Invoke\[[4](https://arxiv.org/html/2605.13391#bib.bib11)\]further studied tool invocation rewriting for zero\-shot tool retrieval\. Meanwhile, ToolGen\[[41](https://arxiv.org/html/2605.13391#bib.bib23)\]framed tool selection as a generative decision\-making problem, introducing specific tool tokens into the model’s vocabulary to achieve unified retrieval and invocation\. While these approaches partially relieve context load and significantly boost selection efficiency, they still fundamentally rely on one\-shot retrieval\. Consequently, they remain prone to overlooking critical tools required for subsequent steps in complex, long\-horizon tasks\.

Building on these foundations, to address multi\-step tool scheduling and dependency issues in complex tasks, researchers have begun exploring more structured and dynamic paradigms for tool organization and retrieval\. One representative line of work attempts to organize the tool space by explicitly modeling the execution logic and input\-output dependencies between tools\. For instance, Voyager\[[38](https://arxiv.org/html/2605.13391#bib.bib37)\], demonstrated that LLM agents can autonomously build growing skill libraries through lifelong learning in open\-ended environments, providing an early precedent for hierarchical skill organization\. ToolNet\[[25](https://arxiv.org/html/2605.13391#bib.bib24)\]moves away from the traditional flat tool list, instead constructing a structured tool network based on state transitions and invocation paths\. Through explicit routing mechanisms, it guides the agent to explore the tool space sequentially, thereby enhancing selection efficiency and robustness in large\-scale environments\. Another approach deeply couples tool retrieval with the reasoning process\. For example, ToolReAGt\[[3](https://arxiv.org/html/2605.13391#bib.bib10)\]proposes a reasoning\-integrated retrieval\-augmented mechanism that combines the agent’s intermediate reasoning results with the current context state to dynamically retrieve the most relevant toolsets for subsequent operations, enabling continuous adjustment of the tool selection process\. Furthermore, recent studies have begun abstracting tools into reusable skills, building systematic capability organization mechanisms around large\-scale skill libraries\. Graph of Skills\[[17](https://arxiv.org/html/2605.13391#bib.bib25)\]proposes a dependency\-aware structural retrieval method for large\-scale agent skill libraries\. Its core idea is to construct executable skill graphs during the offline phase and, during inference, combine semantic\-lexical seed retrieval, Personalized PageRank, and context budget constraints to dynamically recall skill sets that include critical prerequisites\. This mitigates the issue where pure semantic retrieval might overlook necessary prerequisite skills in complex tasks\. In contrast, SkillNet\[[20](https://arxiv.org/html/2605.13391#bib.bib26)\]focuses on the unified management of agent skills from an infrastructure perspective\. By building an open skill network that includes mechanisms for skill creation, evaluation, organization, and connection, it transforms scattered execution experiences into reusable, composable, and evaluable skill assets, supporting long\-term capability accumulation across tasks\. These studies indicate that agent capability organization is evolving from tool\-centric retrieval toward structured management mechanisms oriented around skill dependencies, capability reuse, and long\-term evolution\.

Although these approaches improve scalability in large\-scale tool spaces from perspectives such as semantic retrieval, structured organization, and dynamic code generation, they still fundamentally model tool acquisition as a problem of one\-shot retrieval or on\-demand generation, lacking an explicit mechanism to capture the evolving nature of tool requirements during task execution\. In long\-horizon RS tasks, agents often need to continuously adjust subsequent tool selections based on intermediate states, making tool acquisition inherently closer to a process of active exploration paradigm\. Motivated by this observation, this paper further formulates tool selection as an active exploration process within a structured tool space, enabling agents to incrementally discover and acquire tools during reasoning rather than making a static decision in a single step\.

## IIIMethod

This paper models the task\-solving process of the RS agent as a sequential decision\-making problem within a structured tool space, where information unfolding, tool discovery, tool execution, and result generation are uniformly integrated into a single policy space\. Distinct from the traditional passive tool selection paradigm, this paper defines the agent as an active tool explorer\. By explicitly modeling the dynamic evolution of the currently visible tool information set, the agent is empowered to autonomously determine when to expand the visible tool space and which specific subset of tool information to explore based on the current task, thereby advancing subsequent reasoning and execution\. Building upon this core modeling, we further construct a hierarchical skill tree and design a corresponding progressive disclosure strategy\. The overall methodology of RS\-Claw comprises three main components: unified sequential decision\-making modeling, hierarchical skill tree construction, and progressive disclosure strategy design, as illustrated in Fig\.[2](https://arxiv.org/html/2605.13391#S3.F2)\.

![Refer to caption](https://arxiv.org/html/2605.13391v1/fig2.png)Figure 2:The overall framework of the progressive active tool exploration mechanism based on a hierarchical skill tree\. The top panel illustrates the unified sequential decision\-making modeling, where the agent autonomously selects actions from the defined action space𝒜\\mathcal\{A\}based on the current decision statehkh\_\{k\}\. The bottom panel details the progressive disclosure strategy along the skill tree\. Driven by the policy, the agent proactively executes exploration actions to sequentially acquire tool briefs and detailed execution documents, thereby progressively accumulating context observationsokinfoo\_\{k\}^\{\\mathrm\{info\}\}and dynamically expanding the callable tool set𝒱k\\mathcal\{V\}\_\{k\}\.### III\-AUnified Sequential Decision\-Making Modeling

Consider a user queryxxand a large\-scale RS tool library𝒯=\{t1,…,tN\}\\mathcal\{T\}=\\\{t\_\{1\},\\dots,t\_\{N\}\\\}, where each tooltit\_\{i\}corresponds to a semantic descriptiondid\_\{i\}used to describe its functionality, inputs, and outputs\. Let𝒟=\{d1,…,dN\}\\mathcal\{D\}=\\\{d\_\{1\},\\dots,d\_\{N\}\\\}be the set of all full tool descriptions\. RS tasks typically have long\-horizon dependencies and heterogeneous intermediate representations, and their subsequent decisions depend on the intermediate observations generated by preceding steps\. Therefore, the solving of a user queryxxis essentially not a single\-step matching problem, but a sequential decision\-making problem that continuously evolves with intermediate observations\.

Therefore, the task\-solving process of the RS agent can be formulated as a Partially Observable Markov Decision Process \(POMDP\)\[[12](https://arxiv.org/html/2605.13391#bib.bib44)\]:

ℳ=\(𝒵,𝒜,𝒪,P\)\\mathcal\{M\}=\(\\mathcal\{Z\},\\mathcal\{A\},\\mathcal\{O\},P\)\(1\)where𝒵\\mathcal\{Z\}denotes the latent environmental state space, including the real RS data states and the environmental states generated after tool execution;𝒜\\mathcal\{A\}denotes the agent’s action space;𝒪\\mathcal\{O\}denotes the observation space;P​\(zk\+1∣zk,ak\)P\(z\_\{k\+1\}\\mid z\_\{k\},a\_\{k\}\)denotes the state transition probability\. Letak∈𝒜a\_\{k\}\\in\\mathcal\{A\}denote the specific action taken by the agent when making a decision at thekk\-th step, andok∈𝒪o\_\{k\}\\in\\mathcal\{O\}denote the observation returned by the environment after the action is executed\. Since the latent environmental statezkz\_\{k\}is invisible to the agent, the agent can only make decisions based on the execution history, the currently callable tool set, and the user query\. To this end, we define the unified decision state of the agent at thekk\-th step as:

hk=\(x,𝒱k,τk\)h\_\{k\}=\(x,\\mathcal\{V\}\_\{k\},\\tau\_\{k\}\)\(2\)whereτk=\(o0,a1,o1,a2,o2,…,ak−1,ok−1\)\\tau\_\{k\}=\(o\_\{0\},a\_\{1\},o\_\{1\},a\_\{2\},o\_\{2\},\\dots,a\_\{k\-1\},o\_\{k\-1\}\)represents the interaction history before the agent makes the decision at thekk\-th step\. Specifically, we leto0o\_\{0\}denote the initial context available to the agent before the first decision, while subsequent observations record the results actually observed during the reasoning process\.𝒱k⊆𝒯\\mathcal\{V\}\_\{k\}\\subseteq\\mathcal\{T\}represents the set of tools currently directly callable by the agent at thekk\-th step\. A tool can be called if and only if the semantic descriptiondid\_\{i\}of the tooltit\_\{i\}exists in the current interaction historyτk\\tau\_\{k\}\. Therefore, the callable tool set is defined as:

𝒱k=\{ti∈𝒯∣di∈τk\}\\mathcal\{V\}\_\{k\}=\\\{t\_\{i\}\\in\\mathcal\{T\}\\mid d\_\{i\}\\in\\tau\_\{k\}\\\}\(3\)
The agent selects actionaka\_\{k\}according to the policyπ​\(ak∣hk\)\\pi\(a\_\{k\}\\mid h\_\{k\}\)\. After the actionaka\_\{k\}is executed, the environment transitions to the next latent state and returns a new observationoko\_\{k\}\.

Based on the unified model established above, existing tool acquisition mechanisms can be reformulated as specifically constrained forms within this sequential decision\-making framework\. Among them, the “Flat” strategy passively and statically injects the full tool descriptions𝒟\\mathcal\{D\}into the context\. Its decision state can be written as:

hkflat=\(x,𝒱k,τk\),where​o0=𝒟,𝒱k=𝒯h\_\{k\}^\{\\mathrm\{flat\}\}=\(x,\\mathcal\{V\}\_\{k\},\\tau\_\{k\}\),\\quad\\text\{where \}o\_\{0\}=\\mathcal\{D\},\\ \\mathcal\{V\}\_\{k\}=\\mathcal\{T\}\(4\)This formulation implies that throughout the entire decision\-making process, all tools are defaulted to be directly callable, compelling the agent to make every decision under the interference of massive and voluminous tool descriptions\. Consequently, its selectable actions at thekk\-th step can be written as:

akflat∈\{call​\(ti,θi\),answer​\(y\)\},ti∈𝒱ka\_\{k\}^\{\\mathrm\{flat\}\}\\in\\\{\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\),\\mathrm\{answer\}\(y\)\\\},\\quad t\_\{i\}\\in\\mathcal\{V\}\_\{k\}\(5\)wherecall​\(ti,θi\)\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\)denotes invoking the tooltit\_\{i\}with parametersθi\\theta\_\{i\}under the condition thattit\_\{i\}belongs to the current callable set𝒱k\\mathcal\{V\}\_\{k\};answer​\(y\)\\mathrm\{answer\}\(y\)denotes outputting the final answer and terminating the reasoning\. At this time, its observation information is solely manifested as the tool execution resultokexeco\_\{k\}^\{\\mathrm\{exec\}\}\.

Similarly, the external retrieval paradigm based on the “RAG” strategy can also be viewed as a special case of the aforementioned decision\-making process\. This paradigm recalls a candidate subset from the full tool descriptions𝒟\\mathcal\{D\}via external algorithms, with the update rules falling completely outside the agent’s action space\. Its state also follows the unified format, but both the initial context and the callable tools are passively determined by the external retrieval results:

hkRAG\\displaystyle h\_\{k\}^\{\\mathrm\{RAG\}\}=\(x,𝒱k,τk\),\\displaystyle=\(x,\\mathcal\{V\}\_\{k\},\\tau\_\{k\}\),\(6\)where​o0\\displaystyle\\text\{where \}o\_\{0\}=fsim​\(x,𝒟\),𝒱k=\{ti∣di∈o0\}\\displaystyle=f\_\{\\mathrm\{sim\}\}\(x,\\mathcal\{D\}\),\\quad\\mathcal\{V\}\_\{k\}=\\\{t\_\{i\}\\mid d\_\{i\}\\in o\_\{0\}\\\}wherefsimf\_\{\\mathrm\{sim\}\}denotes the similarity retrieval function independent of the agent’s action space\. At this time, the initial contexto0o\_\{0\}and the callable tool set𝒱k\\mathcal\{V\}\_\{k\}are both generated by the external retriever, and the agent cannot proactively expand this set through its own actions\. Its available actions at thekk\-th step is the same as the flat paradigm:

akRAG∈\{call​\(ti,θi\),answer​\(y\)\},ti∈𝒱ka\_\{k\}^\{\\mathrm\{RAG\}\}\\in\\\{\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\),\\mathrm\{answer\}\(y\)\\\},\\quad t\_\{i\}\\in\\mathcal\{V\}\_\{k\}\(7\)Its observation information is likewise primarily manifested as the tool execution resultokexeco\_\{k\}^\{\\mathrm\{exec\}\}\.

Through the above formulations, it becomes unequivocally clear that existing methods impose static constraints on the update of the callable set𝒱k\\mathcal\{V\}\_\{k\}, forcibly isolating it from the agent’s action space𝒜\\mathcal\{A\}\. Such constraints cause these methods to degenerate into a passive paradigm for tool selection\. Under this passive paradigm, the agent acts merely as a receiver of a predefined toolset, incapable of autonomously guiding the information unfolding process\.

Distinct from the aforementioned methods, this paper fundamentally internalizes the acquisition of tool information as an autonomous decision variable of the agent, proposing an active paradigm for tool exploration\. The decision state of our proposed method can be formulated as:

hkactive=\(x,𝒱k,τk\),where​𝒱k⊆𝒯h\_\{k\}^\{\\mathrm\{active\}\}=\(x,\\mathcal\{V\}\_\{k\},\\tau\_\{k\}\),\\quad\\text\{where \}\\mathcal\{V\}\_\{k\}\\subseteq\\mathcal\{T\}\(8\)Its optional action at thekk\-th step consists of information exploration actions, tool execution actions, and termination actions:

akactive∈𝒜explore∪\{call​\(ti,θi\),answer​\(y\)\}a\_\{k\}^\{\\mathrm\{active\}\}\\in\\mathcal\{A\}\_\{\\mathrm\{explore\}\}\\cup\\\{\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\),\\mathrm\{answer\}\(y\)\\\}\(9\)where𝒜explore\\mathcal\{A\}\_\{\\mathrm\{explore\}\}denotes the tool information acquisition actions that the agent can proactively perform\. Its single\-step observation output satisfies:

ok=\{okinfo,ak∈𝒜explore,okexec,ak=call​\(ti,θi\)\.o\_\{k\}=\\begin\{cases\}o\_\{k\}^\{\\mathrm\{info\}\},&a\_\{k\}\\in\\mathcal\{A\}\_\{\\mathrm\{explore\}\},\\\\ o\_\{k\}^\{\\mathrm\{exec\}\},&a\_\{k\}=\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\)\.\\end\{cases\}\(10\)Here, the information observationokinfoo\_\{k\}^\{\\mathrm\{info\}\}corresponds to the local tool information acquired by the agent during exploration actions, while the execution observationokexeco\_\{k\}^\{\\mathrm\{exec\}\}corresponds to the results of tool invocations\. In stark contrast to the passive paradigm, the history trajectoryτk\\tau\_\{k\}in our method explicitly accumulates tool\-related observations in tandem with the agent’s active exploration, thereby subjecting the callable set𝒱k\\mathcal\{V\}\_\{k\}to the control of the agent’s policy\. Regarding how to impose a high\-dimensional structured organization upon the massive tool library, as well as how to design𝒜explore\\mathcal\{A\}\_\{\\mathrm\{explore\}\}and the information observationokinfoo\_\{k\}^\{\\mathrm\{info\}\}to support the progressive disclosure of tool information, we will elaborate on these in Sections III\-B and III\-C\.

TABLE I:Comparison of mathematical modeling across different tool acquisition paradigms in the sequential decision\-making process\.
### III\-BHierarchical Skill Tree Construction

As modeled in Section III\-A, the dynamic acquisition of tool information is abstracted as the joint evolution of the history trajectoryτk\\tau\_\{k\}and the callable set𝒱k\\mathcal\{V\}\_\{k\}\. However, under the traditional flat organizational structure, when facing the full tool descriptions𝒟\\mathcal\{D\}, the agent lacks explicit search boundaries and effective observation targets\. To address this, inspired by the top\-down hierarchical cognitive logic of humans—from macroscopic intentions to microscopic actions—and hierarchical abstraction in sequential decision\-making\[[37](https://arxiv.org/html/2605.13391#bib.bib45)\], this paper introduces a hierarchical skill tree topological structure\. This imposes an organizational framework upon the high\-dimensional tool space and provides priors for subsequent local information unfolding\. Specifically, let𝒮=\{s1,…,sM\}\\mathcal\{S\}=\\\{s\_\{1\},\\dots,s\_\{M\}\\\}be the set of skill nodes\. We perform an orthogonal partition on the full tool set𝒯\\mathcal\{T\}, mapping it precisely into theseMMskill nodes to form mutually exclusive local tool subspaces𝒯m\\mathcal\{T\}\_\{m\}\. This strict constraint ensures that any underlying RS toolti∈𝒯t\_\{i\}\\in\\mathcal\{T\}uniquely belongs to a single parent skill nodesms\_\{m\}\. Building upon this, we rigorously decouple the massive tool space into three logically progressive information subspaces, which jointly constitute the complete hierarchical skill tree spaceℐ\\mathcal\{I\}to guide the agent in top\-down sequential exploration:

1. 1\.Skill Summary Layer:This layer consists ofMMhigh\-dimensional skill node summariesσ​\(sm\)\\sigma\(s\_\{m\}\), primarily serving to provide a macroscopic functional index for the entire toolset\. Specifically, we employ a concise natural language segment asσ​\(sm\)\\sigma\(s\_\{m\}\)to demarcate the core functional boundary of the branch for the agent, thereby guiding the large language model to rapidly accomplish semantic matching between the macroscopic task intention and the local tool cluster\. This global capability view, constructed at an extremely low token cost, serves directly as the initial contexto0o\_\{0\}for the agent during actual reasoning\. It assists the agent in conducting a coarse screening of macroscopic directions in the early decision\-making stages, effectively preventing it from deviating into completely irrelevant domain branches, and rapidly narrowing the search scope down to a specific capability subspace\.
2. 2\.Tool Catalog Layer:Once the search scope is narrowed, the agent encounters multiple candidate tools within a specific branch\. This layer facilitates local disambiguation by providing brief descriptionsψ​\(ti\)\\psi\(t\_\{i\}\)of these tools\. Given that tools within a specific skill subspace often exhibit high functional similarity, this layer deliberately obscures underlying code\-level parameter specifications to guarantee decision precision, retaining only core functional descriptions, applicability boundaries, and high\-level input/output dependencies\. This filtering mechanism enables the agent to compare similar tools free from the interference of underlying implementation details\. Only after the target tool is definitively selected does the decision process advance to the next stage\.
3. 3\.Tool Documentation Layer:Functioning as the interface for the agent to interact with the environment, this layer is situated at the terminal leaves of the tree topology\. It encompasses the complete semantic descriptionsdid\_\{i\}of the underlying tools, covering detailed execution documents, strict API signatures, and parameter specifications to support the ultimate tool invocation actions\. These machine\-readable execution specifications typically contain numerous cumbersome constraints, which not only cause token overhead surges but also easily induce attention defocus in LLMs over long contexts\. Implementing lazy loading for these detailed pieces of information ensures that the agent only loads and bears this information payload on\-demand after fully clarifying its invocation intention, thereby maximizing the stability of long\-horizon reasoning and the legality of tool invocations\.

This three\-layer architecture transcends a mere segmentation of original tools to represent a profound alignment with the RS task\-solving logic\. Relying on the structured organization between adjacent layers, this architecture strictly adheres to the progressive search path of first determining the capability scope, then screening specific tools, and finally loading execution parameters, enabling the originally unstructured, flat tool set to evolve into a well\-structured tree topology\. During the agent’s top\-down, layer\-by\-layer reasoning process, the exploration granularity is continuously refined, while irrelevant semantic noise is physically isolated and progressively filtered out, thus maintaining a pure context environment for the large language model when executing long\-horizon tasks\.

### III\-CProgressive Disclosure Strategy Design

Based on the static hierarchical skill tree spaceℐ\\mathcal\{I\}constructed in Section III\-B, this section focuses on designing the agent’s dynamic exploration mechanism over this space, directly instantiating the unfolding logic of local tool information into the agent’s policy itself\. Initially, the system only provides the topmost skill layer summaries to the agent as the initial context, and no underlying tools can be directly called yet, i\.e\.:

o0=\{σ​\(sm\)\}m=1M,𝒱1=∅o\_\{0\}=\\\{\\sigma\(s\_\{m\}\)\\\}\_\{m=1\}^\{M\},\\qquad\\mathcal\{V\}\_\{1\}=\\emptyset\(11\)
At thekk\-th step, the agent selects an action based on the current decision statehkh\_\{k\}\. Following the unified definition in Section III\-A, the information exploration actions𝒜explore\\mathcal\{A\}\_\{\\mathrm\{explore\}\}in our method specifically comprise two categories: one isskill​\(sm\)\\mathrm\{skill\}\(s\_\{m\}\), which expands the brief descriptionsψ​\(t\)\\psi\(t\)of all candidate tools under the skill nodesms\_\{m\}; the other isdoc​\(ti\)\\mathrm\{doc\}\(t\_\{i\}\), which expands the detailed documentdid\_\{i\}of the tooltit\_\{i\}\. The observations returned by the information exploration actions will simultaneously affect the history trajectory and the subsequent tool calling permissions\.

When the agent executesak=skill​\(sm\)a\_\{k\}=\\mathrm\{skill\}\(s\_\{m\}\), the system only returns the brief descriptions of all tools under that branch,okinfo=\{ψ​\(t\)∣t∈𝒯m\}o\_\{k\}^\{\\mathrm\{info\}\}=\\\{\\psi\(t\)\\mid t\\in\\mathcal\{T\}\_\{m\}\\\}\. The role of this observationokinfoo\_\{k\}^\{\\mathrm\{info\}\}is to help the agent narrow down the search scope to precisely locate the required tool, but it will not directly unlock the calling permission for the underlying tools \(i\.e\.,𝒱k\+1=𝒱k\\mathcal\{V\}\_\{k\+1\}=\\mathcal\{V\}\_\{k\}\)\. Only when the agent further executesak=doc​\(ti\)a\_\{k\}=\\mathrm\{doc\}\(t\_\{i\}\)will the system return the detailed execution documentdid\_\{i\}of the tooltit\_\{i\}\. At this time, not only will this document be written into the history trajectory, but the underlying tooltit\_\{i\}will also be formally incorporated into the callable set in the next step \(i\.e\.,𝒱k\+1=𝒱k∪\{ti\}\\mathcal\{V\}\_\{k\+1\}=\\mathcal\{V\}\_\{k\}\\cup\\\{t\_\{i\}\\\}\)\. Furthermore, for regular execution actionsak=call​\(ti,θi\)a\_\{k\}=\\mathrm\{call\}\(t\_\{i\},\\theta\_\{i\}\)or the termination actionak=answer​\(y\)a\_\{k\}=\\mathrm\{answer\}\(y\), the environment solely returns the physical execution results or a termination flag, without introducing any new available tools \(i\.e\.,𝒱k\+1=𝒱k\\mathcal\{V\}\_\{k\+1\}=\\mathcal\{V\}\_\{k\}\)\.

This update rule indicates that the skill layer and the tool catalog layer return intermediate information observations, which are written into the history trajectory and influence subsequent reasoning, but do not immediately make the underlying tools callable\. Only after the agent proactively requests the detailed document of a specific tool is that documentdid\_\{i\}incorporated into the history trajectory, and the corresponding tooltit\_\{i\}enters the callable set𝒱k\+1\\mathcal\{V\}\_\{k\+1\}\. Therefore, tool discovery is no longer manifested as a one\-time statically injected context, but rather a sequential process autonomously advanced by the agent\.

Because only local information relevant to the current task is unfolded at each step, our method compresses the context load from a global𝒪​\(N\)\\mathcal\{O\}\(N\)information overhead down to a local𝒪​\(K\)\\mathcal\{O\}\(K\)information overhead, whereK≪NK\\ll Nand is dynamically determined by the policy\. More importantly, this mechanism can reduce the influx of irrelevant information into the context, thereby alleviating context pollution and token consumption caused by full\-set registration\. Consequently, the abstract active tool exploration process proposed in Section III\-A is concretely implemented at runtime as a policy that progressively accumulates tool\-related observations along the hierarchical skill tree and synchronously expands the callable tool set𝒱k\\mathcal\{V\}\_\{k\}\.

## IVExperiments

### IV\-AExperimental Setup

Benchmark\.We evaluate RS\-Claw on the Earth\-Bench, which comprises 248 RS agent analysis questions spanning multiple RS tasks, including spectral index computation, time\-series analysis, change detection, building and object detection, and multi\-source data fusion\. The benchmark defines two evaluation modes: Autonomous Planning mode \(AP\), corresponding to implicit step specification, which assesses the agent’s ability to autonomously plan solution paths—the agent must independently decide which tools to invoke and in what order; and Instruction Following mode \(IF\), corresponding to explicit step specification, which assesses the agent’s ability to translate human instructions into executable actions—reasoning steps are provided externally, and the agent is responsible for converting them into correct tool calls\. Since the official results were obtained in an environment where the ChangeOS dependency was available, whereas the released benchmark does not provide the required deep learning model or precomputed outputs for this tool, we exclude 14 ChangeOS\-related questions \(IDs 216–225 and 242–245\) from all reported results for a fair comparison; consequently, all metrics are computed over 234 questions\.

Baselines\.We compare RS\-Claw, which actively explores the tool space, against two passive\-selection baseline agents:

1. 1\.Flat, the official Earth\-Agent baseline, which writes all available tools as an unstructured flat list into the system prompt, from which the agent directly selects tools; and
2. 2\.RAG, a retrieval\-augmented tool\-aware agent that, before each question, retrieves the 19 most relevant tools from the tool library based on question semantics \(with the file enumeration toolget\_filelistadditionally forced\-included\), and provides the retrieval results as tool context for the agent’s planning and invocation\.

Both baselines and RS\-Claw are built on the ReAct framework, where agents complete multi\-step tool\-call reasoning through alternating Thought–Action–Observation cycles\.

Models\.To assess the generalizability of our method across different model capability levels, we conduct experiments on three large language models: GPT\-5, DeepSeek\-V3\.1, and Qwen3\-32b\. Ablation studies are conducted exclusively on Qwen3\-32b to control experimental cost\.

Skill tree configuration\.RS\-Claw organizes tools into five skill categories consistent with the Earth\-Agent official tool taxonomy: Index, Inversion, Perception, Analysis, and Statistics\. Full tool listings and counts for each skill are provided in Appendix[B\-B](https://arxiv.org/html/2605.13391#A2.SS2)\.

Evaluation metrics\.We evaluate each method along two dimensions: accuracy and context overhead\. For accuracy, we adopt the two\-tier evaluation protocol of the Earth\-Bench\. End\-to\-end metrics include Accuracy and Efficiency; step\-level metrics include Tool\-Any\-Order, Tool\-In\-Order, Tool\-Exact\-Match, and Parameters\. The main tables report the three core metrics—Accuracy, Tool\-Any\-Order, and Tool\-In\-Order—with complete results provided in Appendix[A](https://arxiv.org/html/2605.13391#A1)\. For context overhead, we measure the impact of different tool disclosure strategies on context length using the average number of input tokens per question and per turn\.

### IV\-BExperimental Results

TABLE II:Accuracy metrics comparison\. Accuracy metrics of RS\-Claw \(active tool exploration\) against two passive tool selection baselines \(Flat, RAG\) across three models and two evaluation modes \(AP and IF\)\. Complete metrics are provided in Appendix Table[VI](https://arxiv.org/html/2605.13391#A1.T6)\.TABLE III:Context overhead metrics comparison\. Average input token counts of RS\-Claw \(active tool exploration\) against two passive tool selection baselines \(Flat, RAG\) across three models and two evaluation modes \(AP and IF\), measured in tokens/question and tokens/turn\.Accuracy analysis\.As shown in Table[II](https://arxiv.org/html/2605.13391#S4.T2), RS\-Claw achieves higher accuracy than both passive baselines across all three models and both evaluation modes, with tool\-matching metrics also leading comprehensively, validating the effectiveness of the progressive skill tree disclosure strategy\. Notably, the improvement margin increases as model capability decreases: in AP mode on Qwen3\-32b, RS\-Claw outperforms Flat by 12\.45 percentage points, far exceeding the 3\.00\-point gain on GPT\-5\. This indicates that progressive disclosure effectively alleviates constrained reasoning space and attention defocus across models of varying capability levels, with models more sensitive to context load demonstrating larger gains\. RAG achieves lower accuracy than RS\-Claw on all three models, with weaker step\-level metrics as well\. As a passive tool recipient, the RAG agent delegates tool selection to an external embedding model rather than the agent itself; the LLM cannot actively adjust the tool scope based on intermediate results emerging during reasoning, making it difficult to cover the critical tools required by subsequent reasoning steps in long\-horizon tasks\. This constitutes an inherent limitation of insufficient tool coverage in multi\-step RS analysis tasks\. The two passive baselines thus fail on opposite ends of the trade\-off—context overload versus incomplete tool coverage—while RS\-Claw’s active exploration mechanism achieves an effective balance on both dimensions\.

Context overhead analysis\.As shown in Table[III](https://arxiv.org/html/2605.13391#S4.T3), RS\-Claw substantially reduces the total input tokens per question compared to Flat—in AP mode on Qwen3\-32b, the input token compression ratio reaches approximately 86%, with tokens per turn reduced by 81%, directly validating the core design objective of progressive disclosure: on\-demand tool loading that dramatically reduces context overhead\. Consistent with the accuracy analysis, the compression ratio also varies with model capability: GPT\-5 achieves a compression ratio of approximately 20%, far lower than that of Qwen3\-32b, indicating that the context savings from progressive disclosure differ across models, with models more sensitive to context load benefiting more substantially\. RS\-Claw incurs lower tokens per turn than RAG overall, owing to the fact that RS\-Claw loads only the currently relevant nodes per turn, whereas RAG carries a fixed number of tool descriptions every turn\. Notably, the total per\-question cost advantage of RS\-Claw over RAG varies across models: on Qwen3\-32b, RS\-Claw incurs lower total overhead than RAG, whereas on GPT\-5, RS\-Claw incurs higher total overhead than RAG—a discrepancy attributable to differences in the depth of skill tree exploration across models\. Taken together, the two passive baselines occupy opposite ends of the context\-overhead spectrum—Flat trades efficiency for completeness, while RAG trades tool coverage for lower overhead—whereas RS\-Claw’s active exploration mechanism keeps context consumption locally bounded while preserving tool coverage, achieving an effective balance between the two\.

### IV\-CAblation Study

In this section, we validate the necessity of two core design dimensions in RS\-Claw’s progressive disclosure strategy through two sets of controlled variants: the semantic organizational structure of the skill tree, and the progressive disclosure mechanism that retains the skill summary layer within the three\-tier skill tree\. We compare the following three variants on Qwen3\-32b: RS\-Claw \(the agent used in the main experiments, employing semantically grouped three\-tier skill tree and two\-step progressive disclosure\); Random \(retaining the three\-tier skill tree and two\-step progressive disclosure but randomly assigning tools to five skill nodes, disrupting the semantic coherence of the skill summary layer\); and 2layers \(removing the skill summary layer, retaining only the tool catalog layer and tool documentation layer, and embedding all tool names and brief descriptions by category directly into the system prompt, from which the agent selects target tools and then calls doc to retrieve detailed execution documents—degenerating to a one\-step disclosure containing only the doc information exploration action without the skill action\. This represents an intermediate state between the flat Flat paradigm and RS\-Claw\)\. Three variants form an ordered gradient in tool disclosure structure: Flat injects all tool descriptions into the context at once; 2layers retains full tool\-name visibility but loads detailed execution documents on demand, essentially degenerating the three\-tier skill tree of Section III\-B to two tiers; RS\-Claw fully internalizes the progressive unfolding of tool information into the agent’s sequential decision\-making through the complete three\-tier structure\. Complete accuracy metrics are provided in Appendix Table[VII](https://arxiv.org/html/2605.13391#A1.T7)\.

TABLE IV:Ablation accuracy metrics for progressive disclosure strategy\. Accuracy metrics of the two ablation variants \(Random, 2layers\) compared with RS\-Claw under AP and IF modes\.TABLE V:Ablation context overhead for progressive disclosure strategy\. Average input token counts of the two ablation variants \(Random, 2layers\) compared with RS\-Claw under AP and IF modes\.Necessity of semantic organizational structure of the skill tree\.As shown in Tables[IV](https://arxiv.org/html/2605.13391#S4.T4)and[V](https://arxiv.org/html/2605.13391#S4.T5), Random achieves 9\.87 percentage points lower accuracy than RS\-Claw in AP mode, while tokens per question increase by 43%\. When skill groupings are randomized, the semantic prior provided by the skill layer becomes invalid; the agent can no longer effectively narrow the candidate set based on skill summaries and must explore incorrect skill nodes multiple times to locate target tools\. These futile exploration turns accumulate continuously in the context, exacerbating restricted reasoning space and attention defocus, ultimately degrading both accuracy and context overhead\.

Necessity of retaining the skill summary layer in the progressive disclosure mechanism\.As shown in Tables[IV](https://arxiv.org/html/2605.13391#S4.T4)and[V](https://arxiv.org/html/2605.13391#S4.T5), 2layers exposes all tool names directly in the system prompt, bypassing the skill information exploration action required to route to the skill summary layer for tool selection\. Its step\-level metrics outperform RS\-Claw, as the improved tool visibility directly enhances tool discovery\. However, under Qwen3\-32b, this gain comes at the cost of reasoning space—a cost directly reflected in the contradiction where 2layers achieves a higher Tool\-Any\-Order \(58\.07\) than RS\-Claw \(50\.22\) yet a lower Accuracy \(25\.75\) versus RS\-Claw \(33\.05\): the improvement in tool discovery does not translate into higher final answer accuracy, indicating that the compression of reasoning space offsets the benefit of improved tool visibility, ultimately yielding lower accuracy\. This directly validates the judgment of Section III\-B from the perspective of hierarchical skill tree design: the skill summary layer is the key barrier controlling context scale—omitting it causes the two\-tier structure to surpass the three\-tier in tool discovery yet fall significantly behind in end\-to\-end accuracy\. Appendix[D\-B](https://arxiv.org/html/2605.13391#A4.SS2)provides two concrete trajectory comparisons that further illustrate this mechanism: both pairs share identical Tool\-Any\-Order and Tool\-In\-Order scores, ruling out any difference in tool discovery ability and attributing failure directly to constrained reasoning space after omitting the skill summary layer, manifesting as intermediate\-file confusion and premature truncation of multi\-step planning\.

Together, the two ablation groups demonstrate that the semantic organizational structure of the skill tree and the progressive disclosure mechanism retaining the skill summary layer are both indispensable\. The former provides the agent with an effective prior for selecting the skill summary layer; the latter ensures that tool information is expanded on demand\. These two mechanisms jointly constrain the tool information entering the context to the locally relevant subset for the current task, thereby avoiding constrained reasoning space\.

### IV\-DTool Scalability Experiments

This section examines the robustness of the progressive disclosure strategy under continuous expansion of the tool library along two dimensions: same\-domain scaling \(starting from the minimum tool set required for the task, progressively adding same\-domain irrelevant tools up to the full set of 104\) and cross\-domain scaling \(injecting cross\-domain tools entirely unrelated to RS tasks into the tool library\)\. All experiments use the Qwen3\-32b model\.

#### IV\-D1Same\-Domain Tool Scaling

The experiment fixes the skill package structure and, starting from GT \(the minimum tool set required to complete tasks, extracted from the Earth\-Bench official ground\-truth trajectories\), progressively injects same\-domain irrelevant tools into the tool library in increments of 20 until all 104 tools are covered \(All Tools, i\.e\., the RS\-Claw configuration in the main experiments\), forming six scale gradients\. The goal is to examine how context overhead and task accuracy respond to increasing total tool count under the two disclosure paradigms\. Complete numerical results are provided in Appendix Tables[VIII](https://arxiv.org/html/2605.13391#A1.T8)and[IX](https://arxiv.org/html/2605.13391#A1.T9)\.

![Refer to caption](https://arxiv.org/html/2605.13391v1/fig3.png)Figure 3:Accuracy and context overhead curves under same\-domain tool scaling\. AP mode accuracy \(left axis\) and per\-question token overhead \(right axis, K\) of RS\-Claw and Flat as the tool library scales from the minimum set \(GT\) to the full toolset \(104 tools\)\.As shown in Fig\.[3](https://arxiv.org/html/2605.13391#S4.F3), when the tool library contains only the minimum required tool set \(GT\), Flat achieves slightly higher accuracy and slightly lower per\-question token overhead than RS\-Claw—under zero\-redundancy conditions, full registration introduces no semantic noise, and direct tool visibility eliminates the extra navigation overhead of the skill layer, allowing Flat’s direct\-visibility advantage to fully manifest\. This comparison indicates that the advantage of RS\-Claw over Flat is not unconditional, but rather stems from the context load introduced by tool scale expansion\. As tool count increases, the gap rapidly reverses: RS\-Claw accuracy remains broadly stable while Flat declines continuously, falling below RS\-Claw after GT\+20 with the gap generally widening thereafter; token overhead likewise diverges markedly, with RS\-Claw growing gradually while Flat expands near\-linearly, reaching a growth of over 1,100% in AP mode\. A similar trend is observed in IF mode\. By constraining visible tool information to a locally bounded scope at all times, progressive disclosure suppresses both context inflation and the performance degradation caused by escalating candidate interference as tool scale grows, whereas Flat deteriorates continuously on both accuracy and overhead dimensions as tool count increases\.

#### IV\-D2Cross\-Domain Tool Scaling

Building on the same\-domain scaling experiments, this section further examines the behavior of the two disclosure paradigms when the tool library is injected with cross\-domain tools entirely unrelated to RS tasks\. Starting from the Earth\-Bench \(104 RS tools\), we sequentially inject API\-Bank\[[19](https://arxiv.org/html/2605.13391#bib.bib34)\]\(75 general\-purpose tools covering account authentication, calendar reminders, financial services, healthcare, etc\., totaling 179 tools, denoted stage1\) and ToolBench\[[29](https://arxiv.org/html/2605.13391#bib.bib7)\]\(55 tools covering advertising, business, music, etc\., further expanding to 234 tools on top of stage1, denoted stage2\)\. All newly added tools provide no practical utility for Earth\-Bench questions, constituting pure cross\-domain semantic noise\. Complete numerical results are provided in Appendix Tables[X](https://arxiv.org/html/2605.13391#A1.T10)and[XI](https://arxiv.org/html/2605.13391#A1.T11)\.

![Refer to caption](https://arxiv.org/html/2605.13391v1/fig4.png)Figure 4:Accuracy and context overhead comparison under cross\-domain tool scaling\. AP mode accuracy \(left\) and per\-turn token overhead \(right, K\) of RS\-Claw and Flat across three stages of cross\-domain tool library expansion \(104→\\to179→\\to234 tools\)\.As shown in Fig\.[4](https://arxiv.org/html/2605.13391#S4.F4), after injecting cross\-domain tools, neither paradigm experiences significant accuracy collapse—RS\-Claw decreases modestly but remains broadly stable, while Flat consistently remains at a lower level\. This differs from the continuous accuracy decline of Flat observed in the same\-domain scaling experiments; a plausible explanation is that cross\-domain tools are semantically distant from RS tasks, making it easier for the model to filter them out during reasoning\. However, on the overhead dimension, the divergence remains pronounced: Flat tokens per turn grow linearly with total tool count, while RS\-Claw remains essentially stable\. This result directly validates the core advantage of progressive disclosure: the on\-demand loading mechanism ensures that descriptions of irrelevant tools never enter the context, so tool library expansion has minimal impact on actual context consumption\. Progressive disclosure thus combines controllable overhead with robust accuracy in open\-world tool environments\.

Taken together, the same\-domain and cross\-domain scaling experiments demonstrate that the progressive disclosure strategy exhibits favorable robustness under both typical tool library expansion scenarios: accuracy degradation is relatively gradual, and context overhead growth is substantially lower than that of Flat\. This property endows the approach with stronger scalability for real\-world applications characterized by dynamic growth of open\-world tool libraries\.

## VDiscussion

### V\-ALimitations

Although RS\-Claw demonstrates consistent improvements over both baselines across multiple models and evaluation modes, several limitations remain\.

First, the five skill categories in the current skill tree are directly inherited from the Earth\-Agent official tool taxonomy, constituting a static, manually predefined topology\. When the tool library undergoes substantial changes or is migrated to a new domain, domain experts are required to redesign the tree structure, and the existing grouping exhibits granularity imbalance across skill nodes\.

Second, progressive disclosure incorporates tool exploration into the agent’s decision space, yet the quality of exploration paths fundamentally depends on the planning capability of the underlying LLM; when the model’s own planning ability is insufficient, the agent may still make suboptimal exploration decisions\.

Third, the skill→\\todoc two\-step exploration introduces additional interaction rounds, and the advantage of progressive disclosure stems from the context load imposed by tool scale—in scenarios with a small\-scale tool library, the extra exploration overhead may outweigh the benefits\.

Finally, this work is validated solely on the Earth\-Bench with only three models; the generalization of the method to more domains and larger\-scale tool libraries warrants further investigation\.

### V\-BConnections with other work

#### V\-B1Relationship to existing RS agent systems

Works such as RS\-Agent, Earth\-Agent, ChangeAgent, and GeoLLM\-Squad have advanced RS tasks from single\-step perception toward the automated execution of complex workflows\. However, these systems share a common implicit assumption: the tool space is fully visible before task execution begins, and the agent plans and invokes tools over a static, predefined toolset\. As the scale of specialized RS tool libraries continues to grow, the context load and declining tool selection efficiency arising from this assumption become increasingly pronounced\. Building upon the ReAct reasoning framework established by these works, RS\-Claw further incorporates tool acquisition itself into the agent’s decision space, allowing the callable tool set to evolve dynamically throughout the reasoning process\. This provides a more scalable tool management mechanism for long\-horizon RS tasks without requiring any modification to the underlying model\.

#### V\-B2Relationship to general\-purpose tool selection and organization methods

Whether retrieval\-based methods such as ToolReAGt, Gorilla, and ToolLLM, or hierarchical organization methods such as AnyTool, ToolNet, and Graph of Skills, all share a common limitation: the triggering logic of tool acquisition is controlled by an external module and remains decoupled from the agent’s reasoning process\. The former relies on an external retriever to lock in the candidate toolset in a single pass before task execution; the latter introduces hierarchical structures, but the unfolding of those hierarchies is likewise driven by preset program logic rather than the agent’s reasoning intent\. The fundamental distinction of RS\-Claw is that tool information acquisition is fully internalized as the agent’s own decision variable\. The exploration path is an output of the policy, and the callable tool set evolves continuously with intermediate reasoning states, enabling the agent to dynamically adjust its tool exploration direction according to actual task demands during long\-horizon tasks\.

### V\-CSignificances

#### V\-C1Significance for the RS domain

RS tasks are characterized by long\-horizon workflow chains, high semantic similarity among tool descriptions, and complex intermediate states — precisely the conditions under which the full tool registration paradigm faces its most severe bottleneck\. RS\-Claw is designed directly against these domain\-specific pain points, validating the effectiveness of the active exploration mechanism within multi\-source RS tool spaces and offering a viable path toward automated execution of long\-horizon RS tasks over large\-scale specialized tool libraries\.

#### V\-C2Practical significance

RS\-Claw operates entirely at the tool\-side architecture level and requires no model fine\-tuning, making it a plug\-and\-play context management strategy that can be directly integrated into any ReAct\-based agent system\. The method achieves substantial reductions in context overhead while maintaining accuracy gains over both baselines, demonstrating that lowering inference cost and preserving task completion quality are not mutually exclusive\. Tool scaling experiments further show that RS\-Claw’s context overhead remains approximately stable as the tool library grows, a property that translates directly into engineering value for real\-world deployment scenarios characterized by dynamically expanding tool libraries\.

#### V\-C3Theoretical significance

This paper unifies passive tool exploration paradigms \(Flat and RAG\) and the active exploration paradigm within a single sequential decision\-making framework, casting both paradigm classes as special cases of the same formalization and providing a unified modeling perspective on the tool selection problem\. The framework explicitly distinguishes between tool information acquisition and tool execution as two distinct types of observations, revealing that tool acquisition is inherently a dynamic process that evolves with the reasoning state rather than a one\-shot static decision\.

## VIConclusion and Future Work

### VI\-AConclusion

This paper addresses the context bottleneck faced by RS agents operating over large\-scale tool libraries, and proposes RS\-Claw, a progressive active tool exploration mechanism based on a hierarchical skill tree\. We internalize tool information acquisition as the agent’s autonomous decision variable, incorporating tool exploration actions into the policy space through unified POMDP\-based sequential decision\-making modeling\. A three\-tier hierarchical skill tree coupled with a progressive disclosure strategy enables on\-demand tool loading, effectively compressing tool context overhead\. Experiments on the Earth\-Bench demonstrate that RS\-Claw outperforms both passive baselines \(Flat and RAG\) across three models and two evaluation modes, with improvement margins increasing as model capability decreases and context compression reaching up to 86%\. Ablation studies confirm the necessity of both the semantic organizational structure and the three\-tier progressive disclosure mechanism, while tool scaling experiments validate the method’s scalability\. As a lightweight, plug\-and\-play tool organization strategy, progressive disclosure offers a new perspective for agent context management in large\-scale tool spaces\.

### VI\-BFuture Work

Situated within the broader OpenClaw\-style paradigm of accomplishing tasks through dynamic tool invocation, RS\-Claw demonstrates the effectiveness of progressive active tool exploration in remote sensing scenarios\. Nevertheless, further advances are still needed to make this tool\-space management paradigm more adaptive, scalable, and broadly applicable\. We identify four promising directions for future research:

1. 1\.Exploring automatic clustering methods based on tool description semantic embeddings to replace manual categorization, and introducing adaptive split\-and\-merge mechanisms that allow the skill tree to automatically adjust its topology and continuously evolve as the tool library grows\.
2. 2\.Constructing exploration trajectory data under the progressive disclosure interaction paradigm and optimizing the model’s exploration strategy within the hierarchical structure through supervised fine\-tuning or reinforcement learning\.
3. 3\.Introducing an exploration budget or dynamic routing mechanism to adaptively adjust the disclosure depth based on task complexity, achieving a better trade\-off between accuracy and overhead\.
4. 4\.Extending the method to more vertical domains, larger\-scale tool libraries, and a broader range of models for validation\.

## Appendix AComplete Experimental Results

This appendix provides complete numerical results for all experiments reported in the main text\. Table[VI](https://arxiv.org/html/2605.13391#A1.T6)lists the full six\-metric accuracy results for the main experiment across all three models\. Table[VII](https://arxiv.org/html/2605.13391#A1.T7)gives complete accuracy metrics for the ablation study\. Tables[VIII](https://arxiv.org/html/2605.13391#A1.T8)and[IX](https://arxiv.org/html/2605.13391#A1.T9)report accuracy and context overhead under same\-domain tool scaling\. Tables[X](https://arxiv.org/html/2605.13391#A1.T10)and[XI](https://arxiv.org/html/2605.13391#A1.T11)report the corresponding results under cross\-domain tool scaling\.

TABLE VI:Complete accuracy metrics for the main experiment\. Six accuracy metrics for RS\-Claw \(active tool exploration\) against two passive tool selection baselines \(Flat, RAG\) across three models and two evaluation modes \(AP and IF\), including Efficiency, Tool\-Exact\-Match, and Parameters not reported in the main text Table[II](https://arxiv.org/html/2605.13391#S4.T2)\.TABLE VII:Complete accuracy metrics for the ablation study\. Six accuracy metrics for the two ablation variants \(Random, 2layers\) and RS\-Claw under AP and IF modes, including Efficiency, Tool\-Exact\-Match, and Parameters not reported in the main text Table[IV](https://arxiv.org/html/2605.13391#S4.T4)\.TABLE VIII:Complete accuracy metrics for same\-domain tool scaling\. Six accuracy metrics for RS\-Claw and Flat across six tool\-count levels \(from GT to all 104 tools\) under AP and IF modes\.TABLE IX:Context overhead metrics for same\-domain tool scaling\. Average input tokens per question and per turn for RS\-Claw and Flat across six tool\-count levels under AP and IF modes\.TABLE X:Complete accuracy metrics for cross\-domain tool scaling\. Six accuracy metrics for RS\-Claw and Flat across three cross\-domain expansion stages \(104→\\to179→\\to234 tools\) under AP and IF modes\.TABLE XI:Context overhead metrics for cross\-domain tool scaling\. Average input tokens per question and per turn for RS\-Claw and Flat across three cross\-domain expansion stages under AP and IF modes\.
## Appendix BSystem Prompt and Skill Tree Design

### B\-ASystem Prompt Template

The following is the complete system prompt used by RS\-Claw\. The\{kit\_table\}placeholder is dynamically populated at runtime with a JSON block listing the five skill kits and their applicable task descriptions \(see Section A\.2\.2\)\.

`RS\-Claw system prompt The \{kit\_table\} block is rendered as: kit\_table rendered value This skill summary constitutes the skill layer \{σ​\(sm\)\}m=1M\\\{\\sigma\(s\_\{m\}\)\\\}\_\{m=1\}^\{M\} of the hierarchical skill tree, providing the agent with coarse\-grained navigation priors before any tool\-level information is loaded\.`

`B\-B Skill Tree Structure The five skill nodes and their constituent tools are listed below\. Each table corresponds to the tool catalog layer \{ψ​\(t\)\}t∈𝒯m\\\{\\psi\(t\)\\\}\_\{t\\in\\mathcal\{T\}\_\{m\}\} returned when the agent calls skill\(kit\)\. Detailed execution documents dtd\_\{t\} are only revealed upon a subsequent doc\(kit\.tool\_name\) call\. Skill Tree: Five Kits and Constituent Tools \(104 tools\) Appendix C Implementation Details C\-A Baseline Implementation Both the Flat and RAG baselines share the same system prompt, which contains only task instructions without any skill\-level navigation information: Flat / RAG baseline system prompt Flat baseline\. The Flat baseline follows the official Earth\-Agent implementation\. All 104 domain\-specific tools are exposed to the agent via the Model Context Protocol \(MCP\)\. At agent initialization, the MCP client connects to the tool servers and loads the complete tool list; each tool’s name and full description are automatically injected into the LLM’s context as part of the function\-calling schema\. The system prompt contains only task instructions with no skill\-level navigation information\. The agent directly selects and invokes tools from the full flat list in each reasoning step\. RAG baseline\. The RAG baseline uses a retrieval\-augmented tool selection strategy\. All 104 tools are loaded from the MCP servers at startup\. A FAISS vector index is built offline over tool descriptions using a local Ollama embedding model \(nomic\-embed\-text\)\. For each question, the query text is embedded and used to retrieve the top\-kk most semantically similar tools \(k=19k=19\)\. The tool get\_filelist is always force\-included regardless of retrieval score, yielding a fixed context of 20 tools per question\. A fresh ReAct agent is instantiated with only the retrieved tool subset for each question\. C\-B Evaluation Metric Computation End\-to\-end metrics\. Accuracy is computed as the fraction of questions for which the agent’s final answer matches the ground\-truth choice\. The agent’s answer is extracted from the response using the <Answer\>X<Answer\> or <Answer\>X</Answer\> tag pattern\. Efficiency is defined as the ratio of the number of tool calls made by the agent to the number of tool calls in the ground\-truth trajectory: Efficiency=Nmodel/NGT\\text\{Efficiency\}=N\_\{\\text\{model\}\}/N\_\{\\text\{GT\}\}\. A value greater than 1 indicates the agent used more tool calls than the reference solution\. Step\-level metrics\. Four step\-level metrics are computed by comparing the agent’s tool call sequence against the ground\-truth trajectory: • Tool\-Any\-Order: the fraction of ground\-truth tools that appear anywhere in the agent’s tool call sequence, regardless of order\. Formally, \|𝒯pred∩𝒯GT\|/\|𝒯GT\|\|\\mathcal\{T\}\_\{\\text\{pred\}\}\\cap\\mathcal\{T\}\_\{\\text\{GT\}\}\|/\|\\mathcal\{T\}\_\{\\text\{GT\}\}\|, where sets are used \(duplicates collapsed\)\. • Tool\-In\-Order: the fraction of ground\-truth tools matched in order, allowing intervening tool calls\. A greedy sequential scan is used: for each expected tool in order, the earliest remaining occurrence in the predicted sequence is consumed\. • Tool\-Exact\-Match: the fraction of positions where the predicted and ground\-truth tool sequences agree exactly, up to the length of the shorter sequence\. Matching stops at the first mismatch\. • Parameters: the fraction of ground\-truth steps for which both the tool name and all input parameters exactly match the prediction\. Matching proceeds step by step and stops at the first mismatch\. All four metrics are soft\-scored \(partial credit per question\) and averaged across questions\. Context overhead metrics\. For each agent turn, the number of input \(prompt\) tokens is recorded from the LLM response metadata\. Two aggregate statistics are reported: 1\. tokens per question—the sum of prompt tokens across all turns for a given question, averaged over all questions; 2\. tokens per turn—the per\-turn prompt token count, averaged over all turns across all questions\. Appendix D Case Study D\-A Comparative Trajectory Cases Against Flat This subsection presents two representative cases to illustrate the effectiveness of the progressive skill tree disclosure strategy at the trajectory level\. The two cases correspond respectively to the two core problems identified in the introduction: tool hallucination \(Case F1\) and restricted reasoning space \(Case F2\)\. In each case, RS\-Claw arrives at the correct answer through actual computation, while the Flat baseline fails or produces a wrong answer on the same task\. D\-A1 Case F1 — ATI Drought Stress Analysis Task: Using satellite thermal bands and surface albedo to compute ATI, calculate the proportion of the Sahel region with ATI <1\.0<1\.0 indicating drought stress in May 2023\. Correct answer: D \(70\.92%\)\. RS\-Claw trajectory \(DeepSeek\-V3\.1, AP\) Flat trajectory \(DeepSeek\-V3\.1, AP\) RS\-Claw navigates two skill layers \(skill\("inversion"\) →\\to skill\("statistics"\)\) to precisely narrow the visible tools to ATI and calculate\_threshold\_ratio, forming a clean two\-step chain and arriving at the correct answer\. The Flat baseline, operating under the full context of 104 tools, is misled after the ATI call by 13 semantically adjacent but functionally irrelevant tools, resulting in repeated failures and ultimately FAIL\. This directly illustrates tool hallucination: injecting the full toolset causes the model to confuse semantically similar tools, while the skill tree’s hierarchical filtering constrains visible tool information to the locally relevant subset\. D\-A2 Case F2 — Dead Sea NDTI Turbidity Change Analysis Task: Using reflectance data of the Dead Sea for August 2020 and 2022, compute NDTI and analyze the turbidity change trend\. Correct answer: B \(change \+0\.66063\+0\.66063, increasing\)\. RS\-Claw trajectory \(GPT\-5, AP\) Flat trajectory \(GPT\-5, AP\) Both agents correctly identify calculate\_batch\_ndti, but their subsequent paths diverge sharply\. RS\-Claw completes the task in 8 steps with 4 distinct tools, maintaining clear intermediate states throughout \(means 121191\.78 / 55128\.48, difference directly matching the answer\)\. The Flat baseline, under the context load of the full toolset, becomes severely overloaded: after the NDTI computation it attempts 11 additional tools, generating a flood of meaningless intermediates \(pixel counts 1226/1114, sum −\-360,000,192, etc\.\) that overwhelm the reasoning, ultimately producing the wrong answer A\. The failure is not a tool selection error but a constrained reasoning space problem: full\-toolset registration prevents the agent from maintaining coherent intermediate states amid the noise—a micro\-level corroboration of the quantitative finding that RS\-Claw incurs substantially lower per\-turn token overhead than Flat\. D\-B Ablation Evidence Cases: RS\-Claw vs\. 2layers In the ablation study, 2layers achieves a higher overall Tool\-Any\-Order \(TAO, 58\.07\) than RS\-Claw \(50\.22\), confirming that pre\-exposing all tool names does improve tool discovery coverage\. Yet 2layers’ accuracy \(25\.75%\) falls below RS\-Claw’s \(33\.05%\)\. In both cases below, Tool\-Any\-Order \(TAO\) and Tool\-In\-Order \(TIO\) scores are identical across the two methods, ruling out any difference in tool discovery ability and attributing failure directly to constrained reasoning space after omitting the skill summary layer: under Qwen3\-32b, greater tool visibility does not translate into higher accuracy—reasoning space is the decisive bottleneck\. D\-B1 Case A1 — Split\-Window LST High\-Temperature Ratio Task: Using thermal Band 31 and 32 data from irrigated farmland in northern Hebei on August 5, 2021, apply the split\-window algorithm to compute LST, then calculate the percentage of high\-temperature pixels \(\>305\>305 K\)\. Correct answer: D \(63\.17%\)\. RS\-Claw trajectory \(Qwen3\-32b, IF, TAO=1\.00, TIO=1\.00\) 2layers trajectory \(Qwen3\-32b, IF, TAO=1\.00, TIO=1\.00\) Both TAO and TIO are 1\.00 for both methods, confirming that 2layers correctly identified all ground\-truth tools—tool visibility is not the issue\. However, after successfully running split\_window to produce lst\_output\.tif, 2layers was drawn to semantically adjacent tools pre\-exposed in the system prompt \(threshold\_segmentation, calculate\_multi\_band\_threshold\_ratio\), repeatedly consulting their docs and calling them, generating a large volume of spurious intermediate steps\. Under the context load of 94,715 tokens, the model ultimately ran calculate\_threshold\_ratio on the raw BT file rather than the LST result, yielding the incorrect 47\.04%\. RS\-Claw confined visible tools to the inversion and statistics scopes via the skill layer, keeping context clean throughout, and completed correct reasoning in just 16 turns and 20,415 tokens\. This case directly demonstrates that constrained reasoning space alone is sufficient to cause the model to confuse intermediate files at the aggregation stage and produce a wrong answer\. D\-B2 Case A2 — MODIS Atmospheric Water Vapor Monthly Mean Task: Using MODIS bands b02, b05, b17, b18, b19 surface reflectance data over the Turpan region in July 2020, estimate daily atmospheric water vapor via the band ratio method and compute the monthly average\. Correct answer: C \(11\.3910\)\. RS\-Claw trajectory \(Qwen3\-32b, AP, TAO=0\.67, TIO=0\.67\) 2layers trajectory \(Qwen3\-32b, AP, TAO=0\.67, TIO=0\.67\) Both TAO and TIO are 0\.67 for both methods, confirming that 2layers correctly identified the two core tools \(band\_ratio and calc\_batch\_image\_mean\)—tool discovery ability is on par with RS\-Claw\. However, in AP mode, the 104 tool names pre\-exposed in 2layers’ system prompt consumed substantial context, leaving the model insufficient reasoning space \(20,992 tokens\) to complete the critical planning step of iterating over all dates in July\. It ran band\_ratio for July 1st only, then immediately passed the single\-day result 9\.615 as the monthly mean, selecting D\. RS\-Claw loaded tool information on demand via the skill layer, preserving ample reasoning space, and autonomously planned the full “per\-date computation →\\to average” pipeline in AP mode, correctly calling band\_ratio for all three dates before averaging to 11\.391\. This case demonstrates that in AP autonomous\-planning mode, insufficient reasoning space directly truncates multi\-step planning depth—the tools were found, but the task was prematurely closed\. References \[1\] \(2026\) Agent skills overview – Claude platform documentation\. Note: \[Online\]\. Available: https://platform\.claude\.com/docs/en/agents\-and\-tools/agent\-skills/overview Cited by: §I\. \[2\] Y\. Bai, X\. Lv, J\. Zhang, H\. Lyu, J\. Tang, Z\. Huang, Z\. Du, X\. Liu, A\. Zeng, L\. Hou, et al\. \(2024\) Longbench: a bilingual, multitask benchmark for long context understanding\. In Proceedings of the 62nd annual meeting of the association for computational linguistics \(volume 1: Long papers\), pp\. 3119–3137\. Cited by: item 1\. \[3\] N\. Braunschweiler, R\. Doddipatla, and T\.\-C\. Zorila \(2025\) ToolReAGt: tool retrieval for LLM\-based complex task solution via retrieval augmented generation\. In Proc\. 3rd Workshop Towards Knowledgeable Foundation Models \(KnowFM\), pp\. 75–83\. Cited by: §I, §II\-B\. \[4\] Y\. Chen, J\. Yoon, D\. S\. Sachan, Q\. Wang, V\. Cohen\-Addad, M\. Bateni, C\.\-Y\. Lee, and T\. Pfister \(2024\) Re\-invoke: tool invocation rewriting for zero\-shot tool retrieval\. arXiv preprint arXiv:2408\.01875\. Cited by: §II\-B\. \[5\] M\. Chi, A\. Plaza, J\. A\. Benediktsson, Z\. Sun, J\. Shen, and Y\. Zhu \(2016\) Big data for remote sensing: challenges and opportunities\. Proceedings of the IEEE 104 \(11\), pp\. 2207–2219\. Cited by: §I\. \[6\] Y\. Du, F\. Wei, and H\. Zhang \(2024\) AnyTool: self\-reflective, hierarchical agents for large\-scale API calls\. arXiv preprint arXiv:2402\.04253\. Cited by: §II\-B\. \[7\] P\. Feng, Z\. Lv, J\. Ye, X\. Wang, X\. Huo, J\. Yu, W\. Xu, W\. Zhang, L\. Bai, C\. He, et al\. \(2025\) Earth\-Agent: unlocking the full landscape of earth observation with agents\. arXiv preprint arXiv:2509\.23141\. Cited by: §I, §II\-A\. \[8\] N\. Gorelick, M\. Hancher, M\. Dixon, S\. Ilyushchenko, D\. Thau, and R\. Moore \(2017\) Google earth engine: planetary\-scale geospatial analysis for everyone\. Remote sensing of Environment 202, pp\. 18–27\. Cited by: §I\. \[9\] M\. Grizonnet, J\. Michel, V\. Poughon, J\. Inglada, M\. Savinaud, and R\. Cresson \(2017\) Orfeo toolbox: open source processing of remote sensing images\. Open Geospatial Data, Software and Standards 2 \(1\), pp\. 15\. Cited by: §I\. \[10\] X\. Guo, J\. Lao, B\. Dang, Y\. Zhang, L\. Yu, L\. Ru, L\. Zhong, Z\. Huang, K\. Wu, D\. Hu, et al\. \(2024\) Skysense: a multi\-modal remote sensing foundation model towards universal interpretation for earth observation imagery\. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp\. 27672–27683\. Cited by: §II\-A\. \[11\] Y\. Hu, J\. Yuan, C\. Wen, X\. Lu, Y\. Liu, and X\. Li \(2025\) RSGPT: a remote sensing vision language model and benchmark\. ISPRS Journal of Photogrammetry and Remote Sensing 224, pp\. 272–286\. Cited by: §II\-A\. \[12\] L\. P\. Kaelbling, M\. L\. Littman, and A\. R\. Cassandra \(1998\) Planning and acting in partially observable stochastic domains\. Artificial intelligence 101 \(1\-2\), pp\. 99–134\. Cited by: §III\-A\. \[13\] K\. Kuckreja, M\. S\. Danish, M\. Naseer, A\. Das, S\. Khan, and F\. S\. Khan \(2024\) GeoChat: grounded large vision\-language model for remote sensing\. In Proc\. IEEE/CVF Conf\. Comput\. Vis\. Pattern Recognit\. \(CVPR\), pp\. 27831–27840\. Cited by: §II\-A\. \[14\] LangChain Team \(2025\-02\) Benchmarking single agent performance\. Note: LangChain Blog\. \[Online\]\. Available: https://blog\.langchain\.com/react\-agent\-benchmarking/\[Accessed: Apr\. 24, 2026\] Cited by: item 1\. \[15\] C\. Lee, V\. Paramanayakam, A\. Karatzas, Y\. Jian, M\. Fore, H\. Liao, F\. Yu, R\. Li, I\. Anagnostopoulos, and D\. Stamoulis \(2025\) Multi\-agent geospatial copilots for remote sensing workflows\. In Proc\. IGARSS 2025, pp\. 1084–1089\. Cited by: §II\-A\. \[16\] P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\.\-T\. Yih, T\. Rocktäschel, et al\. \(2020\) Retrieval\-augmented generation for knowledge\-intensive NLP tasks\. Advances in Neural Information Processing Systems 33, pp\. 9459–9474\. Cited by: §I\. \[17\] D\. Li, Z\. Li, H\. Du, X\. Wu, S\. Gui, Y\. Kuang, and L\. Sun \(2026\) Graph of skills: dependency\-aware structural retrieval for massive agent skills\. arXiv preprint arXiv:2604\.05333\. Cited by: §II\-B\. \[18\] J\. Li, W\. Zhao, J\. Zhao, W\. Zeng, H\. Wu, X\. Wang, R\. Ge, Y\. Cao, Y\. Huang, W\. Liu, et al\. \(2025\) The tool decathlon: benchmarking language agents for diverse, realistic, and long\-horizon task execution\. arXiv preprint arXiv:2510\.25726\. Cited by: §II\-B\. \[19\] M\. Li, Y\. Zhao, B\. Yu, F\. Song, H\. Li, H\. Yu, Z\. Li, F\. Huang, and Y\. Li \(2023\) Api\-bank: a comprehensive benchmark for tool\-augmented llms\. In Proceedings of the 2023 conference on empirical methods in natural language processing, pp\. 3102–3116\. Cited by: §II\-B, §IV\-D2\. \[20\] Y\. Liang, R\. Zhong, H\. Xu, C\. Jiang, Y\. Zhong, R\. Fang, J\.\-C\. Gu, S\. Deng, Y\. Yao, M\. Wang, et al\. \(2026\) SkillNet: create, evaluate, and connect AI skills\. arXiv preprint arXiv:2603\.04448\. Cited by: §II\-B\. \[21\] Y\. Liang, C\. Wu, T\. Song, W\. Wu, Y\. Xia, Y\. Liu, Y\. Ou, S\. Lu, L\. Ji, S\. Mao, et al\. \(2024\) Taskmatrix\. ai: completing tasks by connecting foundation models with millions of apis\. Intelligent Computing 3, pp\. 0063\. Cited by: §II\-B\. \[22\] C\. Liu, K\. Chen, H\. Zhang, Z\. Qi, Z\. Zou, and Z\. Shi \(2024\) Change\-agent: toward interactive comprehensive remote sensing change interpretation and analysis\. IEEE Transactions on Geoscience and Remote Sensing 62, pp\. 1–16\. Cited by: §II\-A\. \[23\] F\. Liu, D\. Chen, Z\. Guan, X\. Zhou, J\. Zhu, Q\. Ye, L\. Fu, and J\. Zhou \(2024\) Remoteclip: a vision language foundation model for remote sensing\. IEEE Transactions on Geoscience and Remote Sensing 62, pp\. 1–16\. Cited by: §II\-A\. \[24\] N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang \(2024\) Lost in the middle: how language models use long contexts\. Transactions of the Association for Computational Linguistics 12, pp\. 157–173\. Cited by: item 2\. \[25\] X\. Liu, Z\. Peng, X\. Yi, X\. Xie, L\. Xiang, Y\. Liu, and D\. Xu \(2024\) ToolNet: connecting large language models with massive tools via tool graph\. arXiv preprint arXiv:2403\.00839\. Cited by: §II\-B\. \[26\] X\. Liu, H\. Yu, H\. Zhang, Y\. Xu, X\. Lei, H\. Lai, Y\. Gu, H\. Ding, K\. Men, K\. Yang, et al\. \(2023\) Agentbench: evaluating llms as agents\. arXiv preprint arXiv:2308\.03688\. Cited by: §II\-B\. \[27\] OpenClaw \(2026\) OpenClaw: open\-source personal AI assistant\. Note: https://github\.com/openclaw/openclawVersion 2026\.3\.8, Accessed: 2026\-03\-09 Cited by: §I\. \[28\] S\. G\. Patil, T\. Zhang, X\. Wang, and J\. E\. Gonzalez \(2024\) Gorilla: large language model connected with massive APIs\. Advances in Neural Information Processing Systems 37, pp\. 126544–126565\. Cited by: item 2, §II\-B\. \[29\] Y\. Qin, S\. Liang, Y\. Ye, K\. Zhu, L\. Yan, Y\. Lu, Y\. Lin, X\. Cong, X\. Tang, B\. Qian, et al\. \(2023\) ToolLLM: facilitating large language models to master 16000\+ real\-world APIs\. arXiv preprint arXiv:2307\.16789\. Cited by: §I, §II\-B, §II\-B, §IV\-D2\. \[30\] C\. Qu, S\. Dai, X\. Wei, H\. Cai, S\. Wang, D\. Yin, J\. Xu, and J\. Wen \(2025\) Tool learning with large language models: a survey\. Frontiers of Computer Science 19 \(8\), pp\. 198343\. Cited by: §II\-B\. \[31\] T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom \(2023\) Toolformer: language models can teach themselves to use tools\. Advances in Neural Information Processing Systems 36, pp\. 68539–68551\. Cited by: §I\. \[32\] A\. Shabbir, M\. A\. Munir, A\. Dudhane, M\. U\. Sheikh, M\. H\. Khan, P\. Fraccaro, J\. B\. Moreno, F\. S\. Khan, and S\. Khan \(2025\) ThinkGeo: evaluating tool\-augmented agents for remote sensing tasks\. arXiv preprint arXiv:2505\.23752\. Cited by: §II\-A\. \[33\] Y\. Shen, K\. Song, X\. Tan, D\. Li, W\. Lu, and Y\. Zhuang \(2023\) HuggingGPT: solving AI tasks with ChatGPT and its friends in Hugging Face\. Advances in Neural Information Processing Systems 36, pp\. 38154–38180\. Cited by: §I\. \[34\] N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\) Reflexion: language agents with verbal reinforcement learning\. Advances in Neural Information Processing Systems 36, pp\. 8634–8652\. Cited by: §II\-A\. \[35\] S\. Singh, M\. Fore, and D\. Stamoulis \(2024\) Evaluating tool\-augmented agents in remote sensing platforms\. arXiv preprint arXiv:2405\.00709\. Cited by: §II\-A\. \[36\] S\. Soni, A\. Dudhane, H\. Debary, M\. Fiaz, M\. A\. Munir, M\. S\. Danish, P\. Fraccaro, C\. D\. Watson, L\. Klein, F\. S\. Khan, et al\. \(2025\) EarthDial: turning multi\-sensory earth observations to interactive dialogues\. In Proc\. Comput\. Vis\. Pattern Recognit\. Conf\. \(CVPR\), pp\. 14303–14313\. Cited by: §II\-A\. \[37\] R\. S\. Sutton, D\. Precup, and S\. Singh \(1999\) Between mdps and semi\-mdps: a framework for temporal abstraction in reinforcement learning\. Artificial intelligence 112 \(1\-2\), pp\. 181–211\. Cited by: §III\-B\. \[38\] G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar \(2023\) Voyager: an open\-ended embodied agent with large language models\. arXiv preprint arXiv:2305\.16291\. Cited by: §II\-B\. \[39\] L\. Wang, W\. Xu, Y\. Lan, Z\. Hu, Y\. Lan, R\. K\.\-W\. Lee, and E\.\-P\. Lim \(2023\) Plan\-and\-solve prompting: improving zero\-shot chain\-of\-thought reasoning by large language models\. arXiv preprint arXiv:2305\.04091\. Cited by: §I\. \[40\] L\. Wang, C\. Ma, X\. Feng, Z\. Zhang, H\. Yang, J\. Zhang, Z\. Chen, J\. Tang, X\. Chen, Y\. Lin, et al\. \(2024\) A survey on large language model based autonomous agents\. Frontiers of Computer Science 18 \(6\), pp\. 186345\. Cited by: §I, §II\-A\. \[41\] R\. Wang, X\. Han, L\. Ji, S\. Wang, T\. Baldwin, and H\. Li \(2024\) ToolGen: unified tool retrieval and calling via generation\. arXiv preprint arXiv:2410\.03439\. Cited by: §II\-B\. \[42\] W\. Xu, Z\. Yu, B\. Mu, Z\. Wei, Y\. Zhang, G\. Li, J\. Wang, and M\. Peng \(2024\) RS\-Agent: automating remote sensing tasks through intelligent agent\. arXiv preprint arXiv:2406\.07089\. Cited by: §I, §II\-A\. \[43\] S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. R\. Narasimhan, and Y\. Cao \(2023\) ReAct: synergizing reasoning and acting in language models\. In Proc\. 11th Int\. Conf\. Learn\. Represent\. \(ICLR\), Cited by: item 1, §II\-A\. \[44\] S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. Griffiths, Y\. Cao, and K\. Narasimhan \(2023\) Tree of thoughts: deliberate problem solving with large language models\. Advances in neural information processing systems 36, pp\. 11809–11822\. Cited by: §II\-A\. \[45\] W\. Zhang, M\. Cai, T\. Zhang, Y\. Zhuang, and X\. Mao \(2024\) EarthGPT: a universal multimodal large language model for multisensor image comprehension in remote sensing domain\. IEEE Transactions on Geoscience and Remote Sensing 62, pp\. 1–20\. Cited by: §II\-A\.`