In real‑world scenarios, agents must combine reasoning with external tools to accomplish complex tasks. Early agents typically interleave reasoning and actions along a single execution chain; as the horizon grows, the chain becomes unwieldy, intermediate dependencies get obscured, and early planning mistakes propagate, reducing reliability.\ \ Recursively breaking a complex task into smaller subtasks is a natural remedy, yet effective decomposition must respect the system's tool capabilities so that each subtask can be handled by the available tools. In practice, tool libraries are often huge; exposing every tool description consumes substantial context, making relevant tool retrieval and clear task boundaries harder.\ \ We propose tool‑aware recursive decomposition, which organizes tools into a hierarchy based on functional relationships. During execution the agent discovers tools on demand and uses the hierarchy to recursively decompose the overall task into a subtask tree, where each level aligns with the capabilities required at that stage. Experiments on several challenging long‑horizon real‑world tasks show up to a 40‑percentage‑point improvement in end‑to‑end success rate over strong baselines. The implementation is open‑source on GitHub: https://github.com/WeiXiang-Mao/TaReD.\ \ Review