面向关键任务基础设施运营的LLM智能体任务感知约束适配
文章背景与核心概要
随着大语言模型（LLM）智能体在关键基础设施（MCI）运营中的广泛应用，其执行任务时所依赖的“约束配置”（Harness，决定了智能体可访问的信息、工具及采取的行动）显得尤为关键。传统系统往往采用“一刀切”的方法，向每个任务暴露功能齐全的完整约束，这不仅造成了资源浪费，还可能影响执行效率。本文首次深入研究了任务感知约束适配（Task-Aware Harness Provisioning），将其建模为任务需求与约束能力之间的资源匹配问题。
为了优化这一过程，作者引入了一种针对MCI任务的新型分类法，通过文献挖掘和控制执行构建了任务到约束的映射，并提出了一种地图引导的升级（Map-Guided Escalation）配置算法。在液体冷却和电网运营等实际场景中的评估表明，约束优化遵循依赖于特定领域的准确率-成本帕累托边界（Pareto frontier），而非单一的通用最优解。这一研究为提升关键基础设施中智能体的高效、安全运行提供了重要的理论与方法支撑。
任务感知约束适配：面向关键基础设施运营的LLM智能体
arXiv: 2608.17433 [cs.AI]
提交时间: 2026年8月18日
作者: Liangtao Lin, Qingang Zhang, Zhaomeng Zhu, Tianwei Zhang, Yonggang Wen
arXiv: 2608.17433 [cs.AI]
Submitted: August 18, 2026
Authors: Liangtao Lin, Qingang Zhang, Zhaomeng Zhu, Tianwei Zhang, Yonggang Wen
执行摘要
大语言模型（LLM）智能体正日益被用于运营关键基础设施（MCI），它们依赖于决定其信息访问权限、工具调用及动作执行的运营“约束（Harness）”。传统系统通常采用千篇一律的方法，向所有任务暴露完全全面的约束，而不考虑实际需求。本文研究了任务感知约束适配，将其构想为任务需求与约束能力之间的资源匹配问题。
为了优化这一过程，作者引入了MCI任务的新型分类法，从文献和受控执行中构建了任务到约束的映射，并提出了一种地图引导升级的适配算法。对液冷和电网运营的评估表明，约束优化遵循依赖于特定领域的准确率-成本帕累托边界，而不是遵循某个通用最优解。
Executive Summary
Large Language Model (LLM) agents are increasingly utilized to operate mission-critical infrastructure (MCI), relying on operational "harnesses" that dictate their access to information, tools, and actions. Traditional systems typically employ a one-size-fits-all approach, exposing a fully comprehensive harness to every task regardless of actual requirements. This paper investigates task-aware harness provisioning, framing it as a resource-matching problem between task demands and harness capabilities.
To optimize this process, the authors introduce a novel classification of MCI tasks, construct task-to-harness mappings from literature and controlled executions, and propose a map-guided escalation provisioning algorithm. Evaluations across liquid cooling and power grid operations reveal that harness optimization follows a domain-dependent accuracy-cost Pareto frontier rather than adhering to a universal optimum.
元数据与分类
- 主要学科: 人工智能 (
cs.AI) - 次要学科: 多智能体系统 (
cs.MA) - 引用格式:
arXiv:2608.17433 [cs.AI] - DOI: 10.48550/arXiv.2608.17433
Metadata & Classification
- Primary Subject: Artificial Intelligence (
cs.AI)- Secondary Subjects: Multiagent Systems (
cs.MA)- Cite As:
arXiv:2608.17433 [cs.AI]- DOI: 10.48550/arXiv.2608.17433
摘要
LLM智能体已被广泛采用来运营关键基础设施（MCI）。这些智能体通常依赖于一个约束，该约束决定了它们可以访问哪些信息、可以使用哪些工具以及可以采取哪些行动。现有系统通常向每个任务暴露相同的全面约束，这可能并非必要并会导致资源浪费。
在本文中，我们专注于识别最优约束配置，并将其视为任务需求与约束提供能力之间的资源匹配问题。为了衡量这种匹配，我们基于底层系统的数学表示对MCI任务进行分类，并根据约束提供的信息数量和类型对约束配置进行排序。然后，我们从两个来源构建任务到约束的映射：挖掘研究文献和测量受控的智能体执行。
利用测得的映射，我们提出了一种新的约束适配算法：地图引导升级（map-guided escalation）。它从特定任务的约束开始，仅在自我检查失败后才扩展到完全配置。
我们在两个代表性的MCI任务中评估了我们的方法： 1. 液冷（Liquid Cooling）： 将智能体准确率从 \(0.652\)（完全配置下）提升至 \(0.715\)，达到了与 Reflexion 相当的准确率，同时节约了 48% 的 Token。 2. 电网（Power Grids）： 完全配置保持了准确率最优，而基于地图的适配提供了成本更低的替代方案。
这些发现表明，约束适配遵循依赖于特定领域的准确率-成本帕累托边界，而不是通用的最优解。
Abstract
LLM agents have been widely adopted to operate mission-critical infrastructure (MCI). These agents normally rely on a harness that determines what information they can access, which tools they can use, and what actions they can take. Existing systems often expose the same comprehensive harness to every task, which may not be necessary and cause resource wastes.
In this paper, we focus on the identification of optimal harness configurations, and view it as a resource-matching problem between what each task requires and what the harness provides. To measure this match, we classify MCI tasks based on the mathematical representation of the underlying system and rank harness configurations by the amount and type of information they provide. We then construct task-to-harness mappings from two sources: mining research literature and measuring controlled agent execution.
Leveraging the measured mapping, we propose a new harness provisioning algorithm: map-guided escalation. It begins with a task-specific harness and expands to full provision only after a failed self-check.
We evaluate our method in two representative MCI tasks: 1. Liquid Cooling: Improves agent accuracy from \(0.652\) (under full provision) to \(0.715\), achieving accuracy comparable to Reflexion while utilizing 48% fewer tokens. 2. Power Grids: Full provision remains accuracy-optimal, while map-based provisioning offers lower-cost alternatives.
These findings show that harness provisioning follows a domain-dependent accuracy-cost Pareto frontier rather than a universal optimum.
访问链接
Access Links