Librarian Agents: RL-trained agents build filesystems that lift RAG accuracy from 44% to 77%

Tinker reports that agents trained with reinforcement learning learn to build new filesystems in anticipation of other agents' queries, raising subsequent RAG accuracy from 44% to 77%.

The work, called Librarian Agents, is described by Tinker as a step toward training agents to work together, and it is supported by the Tinker grants program. The report is a single post on X from Tinker's account, and it does not describe the evaluation setup or the tasks used to measure accuracy.