KAIST's RunningTab Keeps LLM Agents on Task With Environment-Side Requirement Tabs

KAIST and DeepAuto.ai propose RunningTab, a direct workspace interaction framework where the environment — not the model — tracks per-task requirements, read excerpts with provenance, and unopened file candidates. Agents see each requirement beside its best-matching excerpts and get a finish check if requirements remain open. Across three benchmarks and three LLMs, RunningTab consistently beat plain DWI and in-model record baselines.