RunningTab keeps task requirements, file excerpts and unopened candidates outside an AI agent’s context window. The framework links completed requirements to captured content and returns outstanding ones when the agent tries to finish. In a single preprint, tests across three models and three evaluations improve on direct file access and model-written task lists.
RunningTab addresses information that an AI agent reads but leaves out of its finished work. In an initial analysis, approximately one in five required numeric values shown to a model was missing from the delivered document. The framework keeps a task list outside the model’s limited context. The finding comes from a single preprint on file-based knowledge work.[1]
The agent adds requirements while its working environment saves excerpts with file paths and commands. Listed files that remain unopened stay available as candidates. A completed requirement needs a citation to captured content; an unavailable one needs a reason. When the agent first tries to stop, outstanding requirements return for a final check. Matching uses BM25, a method that ranks text by word relevance.[1]
Tests used three models across three benchmark suites, repeating each setting three times. RunningTab scored above direct file access, model-written task lists and final self-review in the selected pairings. Removing the environment’s excerpt and candidate-file capture was associated with the largest drop. Some retained values still failed to reach the final document. The authors advise applying workspace access controls to stored excerpts, which can hold sensitive information.[1]
