Librarian Agents: LLMs Pre-Build the Workspace and Gain 33 Points on BrowseComp-Plus

Researchers have presented Librarian Agents, a framework in which an LLM agent restructures an unstructured set of documents into a hierarchical file system in advance, and a second agent uses it to answer questions. On BrowseComp-Plus, the method outperforms the strongest retrieval and precomputation baselines by 33 percentage points. On the BRIGHT benchmark, on domains the model did not see during training, average accuracy rose by 31 points without any task-specific fine-tuning.

The underlying problem will be familiar to anyone who has built agents on top of large corpora. Classic retrieval is reactive: a query arrives, the system pulls out fragments, and the model tries to assemble an answer from them. When an answer requires connecting facts across several documents, that step becomes the bottleneck. Precomputed structures such as knowledge graphs help only partly: they are built with rigid heuristics that have no idea what questions will be asked later.

The asymmetry behind the idea

The authors start from an observation about modern agent harnesses such as Claude Code. These agents navigate an existing file system well: they open files, follow links, and gather context for the task at hand. But almost nobody trains them to do the opposite, that is, to create the workspace that will be used later.

Librarian Agents fills this gap. The builder agent receives raw documents and produces a hierarchy of synthetic files with links between them. This workspace is then handed to the answering agent. Construction is formulated as a Markov decision process, and the builder's policy is optimized with reinforcement learning. The reward is tied directly to the final answer accuracy of the downstream agent, not to intermediate metrics such as retrieval quality.

What the experiments showed

According to the abstract, the results are as follows:

• BrowseComp-Plus: +33 percentage points over the strongest retrieval and precompute baselines.
• BRIGHT: average accuracy gain of 31 points in zero-shot mode on unseen domains.
• The cost of building the workspace is amortized over repeated queries to the same corpus.
• What gets trained: the builder's policy, not the answering agent.

Transferability is what matters here. If the builder had simply memorized the structure of specific datasets, it would have failed on BRIGHT. A double-digit gain on new domains suggests that RL training produced general techniques for organizing knowledge: how to group material, what to pull into overview files, and where to place cross-references.

The work also has unknowns. The abstract does not give model sizes, the absolute cost of building, or data on how the workspace behaves when the corpus is updated frequently. These are worth looking for in the full text of the paper.

What teams building agents can try

The practical takeaway concerns architecture. If the corpus is relatively stable and queries against it repeat (an internal knowledge base, product documentation, a research archive), preparing the workspace up front pays off better than searching from scratch each time. The build cost is divided by the number of queries, and the more queries there are, the better the economics.

Second: evaluate corpus preparation by final answer accuracy. That is the reward the paper uses, and it filters out structures that look good but are useless. Third: test on held-out domains, otherwise it is easy to mistake overfitting to your own data for quality. For fast-changing corpora, first calculate the cost of rebuilding, because in that scenario the benefit is less clear.