长文本写作代理需要在信息检索、推理和创作之间实现灵活的整合与交互。当前方法依赖于预设的工作流程和僵化的思维模式，在写作前生成提纲，导致写作过程中适应能力受限。本文提出了一种通用的代理框架，通过递归任务分解和三种基本任务类型（即检索、推理和创作）的动态集成，实现了类似人类的适应性写作。我们的方法具有以下特点：1）一种交织递归任务分解与执行的规划机制，消除了对写作流程的人为限制；2）任务类型的集成促进了异构任务分解。在小说创作和技术报告生成的评估中，我们的方法在所有自动评估指标上均优于最先进的方法，这证明了所提出框架的有效性和广泛应用性。

Long-form writing agents require flexible integration and interaction across information retrieval, reasoning, and composition. Current approaches rely on predetermined workflows and rigid thinking patterns to generate outlines before writing, resulting in constrained adaptability during writing. In this paper we propose a general agent framework that achieves human-like adaptive writing through recursive task decomposition and dynamic integration of three fundamental task types, i.e. retrieval, reasoning, and composition. Our methodology features: 1) a planning mechanism that interleaves recursive task decomposition and execution, eliminating artificial restrictions on writing workflow; and 2) integration of task types that facilitates heterogeneous task decomposition. Evaluations on both fiction writing and technical report generation show that our method consistently outperforms state-of-the-art approaches across all automatic evaluation metrics, which demonstrate the effectiveness and broad applicability of our proposed framework.