Much knowledge work involves producing new deliverables from files a workspace already contains, and LLM agents are increasingly being asked to take on that work. A research paper posted on Hugging Face describes direct workspace interaction (DWI), in which an agent searches and reads files directly from a terminal without any indexing.
The paper argues that reaching the files is only half the problem. Nothing tracks what the task asks for, what has been read, and what was listed but never opened, so these details can slip out of the context window without leaving a trace. As a result, an agent may extract a figure and still deliver a report without it.
To address this, the authors present RunningTab, a framework that gives direct workspace interaction an environment-side tab: a per-task record of what the task still owes, kept by the environment alongside the agent. The agent adds its requirements, while the environment records every file read as an excerpt with its provenance and every listed but unopened file as a candidate.
The agent can then see each requirement beside its best-matching excerpts and top unopened candidates, resolve it against matching content, or set it aside with a reason. If the agent tries to finish while requirements remain open, it receives them in a finish check.
The authors validate RunningTab on three benchmarks with three LLMs. They report that it consistently outperforms plain DWI and baselines that keep the record in the model, and that its tab usually holds the values a deliverable needs once seen.