ControlScope: Workflow Revision and Reliability in LLM Agents
The paper tests how far an LLM agent should be allowed to revise a workflow once it is already running.
ControlScope compares keeping the current code, editing only the next tool call’s arguments, and replacing the unfinished workflow from the same execution state. The results are mixed: broader revision can recover more filesystem tasks, but fast draws and AppWorld show only small or inconsistent gains. The authors also find cases where later reviews interrupt viable repairs, and where a cheaper argument-only edit was available but missed. HF Daily Papers' note
ControlScope compares keeping the current code, editing only the next tool call’s arguments, and replacing the unfinished workflow from the same execution state. The results are mixed: broader revision can recover more filesystem tasks, but fast draws and AppWorld show only small or inconsistent gains. The authors also find cases where later reviews interrupt viable repairs, and where a cheaper argument-only edit was available but missed. HF Daily Papers' note
score 4