When an AI agent performs long tasks, like booking a trip or navigating a website, it sometimes misses a step that later steps depend on. Just trying something else often fails because the real issue is that a prerequisite is missing. The difficulty isn't how long the task is, but recovering from these hidden requirements.
The authors propose a method called Verify-Backtrack Loop (VBL) that extends a common agent approach called ReAct. VBL keeps a 'goal stack,' an ordered list of subgoals, and after each action or failure it checks two things: does this subgoal still help the final goal, and is it achievable now? If not, it can insert a missing prerequisite or go back to a previous step.
On three test environments, VBL improved task success over standard ReAct. The biggest gain was on AppWorld, where success rose by up to 22.07 percentage points. Tests showed both checking and diagnosis are necessary; removing either hurts performance.