RunningTab: Direct Workspace Interaction with Environment-Side Tabs
RunningTab keeps the agent’s unfinished obligations outside the model context, beside evidence from the workspace.
The paper frames the problem as direct workspace interaction: agents searching and reading local files without an index while producing deliverables from them. Its claim is that agents lose track of task requirements, opened evidence, and unopened candidates as context fills. RunningTab adds an environment-side record where requirements, excerpts with provenance, and listed-but-unopened files remain visible. In tests on three benchmarks with three LLMs, it outperformed plain DWI and model-side record baselines.
ArXiv · AI/CL/LG's note
The paper frames the problem as direct workspace interaction: agents searching and reading local files without an index while producing deliverables from them. Its claim is that agents lose track of task requirements, opened evidence, and unopened candidates as context fills. RunningTab adds an environment-side record where requirements, excerpts with provenance, and listed-but-unopened files remain visible. In tests on three benchmarks with three LLMs, it outperformed plain DWI and model-side record baselines.
ArXiv · AI/CL/LG's note
score 5