RunningTab: Direct Workspace Interaction with Environment-Side Tabs
RunningTab keeps an external checklist of what an agent still owes while it works through local files.
The paper targets “direct workspace interaction,” where an LLM agent searches and reads workspace files without a prebuilt index. Its problem case is context loss: files get listed, excerpts get seen, and requirements can still disappear before the final deliverable. RunningTab stores task requirements, read excerpts with provenance, and unopened candidate files on the environment side, then checks unresolved requirements before finish. The authors report gains across three benchmarks and three LLMs over plain DWI and model-side recordkeeping. Source: HF Daily Papers' note
The paper targets “direct workspace interaction,” where an LLM agent searches and reads workspace files without a prebuilt index. Its problem case is context loss: files get listed, excerpts get seen, and requirements can still disappear before the final deliverable. RunningTab stores task requirements, read excerpts with provenance, and unopened candidate files on the environment side, then checks unresolved requirements before finish. The authors report gains across three benchmarks and three LLMs over plain DWI and model-side recordkeeping. Source: HF Daily Papers' note
score 4