Librarian Agents introduce a two‑stage framework in which a builder LLM agent proactively reorganizes an unstructured corpus into a hierarchical filesystem of synthetic files and reference links, creating a curated workspace that a downstream answering agent can consult at inference time. The workspace construction process is modeled as a Markov decision process, and the builder’s policy is trained with reinforcement learning to directly maximize the accuracy of the answering agent on future queries. Experiments show that this approach yields a 33‑percentage‑point gain over the strongest retrieval and precomputation baselines on the BrowseComp‑Plus benchmark, while the initial construction cost is amortized across multiple queries. Furthermore, the method exhibits zero‑shot generalization to unseen domains, achieving an average 31‑point accuracy improvement on the BRIGHT evaluation suite. These results demonstrate that anticipatory, RL‑driven curation of long‑context workspaces can substantially enhance the reasoning capabilities of LLM agents over large document collections.