RunningTab: Direct Workspace Interaction with Environment-Side Tabs

RunningTab. Direct Workspace Interaction with Environment-Side Tabs. By Jinheon Baek and four others, from KAIST and DeepAuto AI. Published on the seventh of October.

A great deal of knowledge work consists of producing something new out of files a workspace already holds, and language model agents are beginning to take that work over. Through direct corpus interaction, an agent can search and read any of those files from a terminal, with no indexing step at all. Producing a deliverable from many files in this way is what the authors call direct workspace interaction, or DWI.

Reaching the files, they argue, is only half the task. Nothing keeps track of what the task asks for, what has already been read, and what was listed but never opened. All three slip through the context window without leaving a trace. The failure this produces is very concrete: an agent may extract a figure from a file and still deliver a report that leaves the figure out.

RunningTab equips direct workspace interaction with what the authors call an environment-side tab. It is a per-task record of what the task still owes, and the important word is environment-side. The record is kept by the environment, alongside the agent, rather than living inside the model's own context where things can quietly fall away.

The work of maintaining the tab is divided. The agent contributes one thing: it adds its requirements. The environment does the bookkeeping. Every file the agent reads is recorded as an excerpt together with its provenance, and every file that was listed but never opened is recorded as a candidate.

From there the agent can look at each requirement side by side with its best-matching excerpts and its top unopened candidates. For each requirement it has two honest options. It can resolve it against matching content, or it can set it aside and give a reason. And if the agent tries to finish while requirements are still open, it receives them in a finish check, so an unresolved requirement is put in front of the agent before the deliverable goes out.

RunningTab is validated on three benchmarks with three language models. Across them it consistently outperforms two things: plain direct workspace interaction, and baselines that keep the same kind of record inside the model rather than in the environment. That second comparison is the one that speaks to the design choice, since it holds the idea of a record fixed and changes only where the record lives. The authors also report that the tab usually holds the values a deliverable needs, once those values have been seen.

The contribution is a relocation of memory. An agent working through a large workspace already has the means to reach every file. What it lacks is a place outside its context window where requirements, excerpts and unopened leads persist, and where an open item is raised again before the agent can finish. RunningTab puts that place in the environment rather than in the model.