Language agents tackling long‑horizon tasks in interactive environments often rely on latent preconditions. The main challenge lies not in trajectory length but in recovery: when a prerequisite is missing or an earlier decomposition fails, trying another plausible action rarely helps. ReAct and later reflection or adaptive‑planning methods only support local correction via free‑form text, lacking an explicit control state that links the terminal goal, the active goal, and the failure source. We introduce the Verify--Backtrack Loop (VBL), a training‑free extension of ReAct that turns precondition discovery and failure diagnosis into structured planning‑state updates. VBL keeps a bounded goal stack and, upon goal activation, execution failure, or state change, jointly verifies two horizons—whether the active goal contributes to the terminal goal and whether it is feasible in the current state. These dual‑horizon judgments, together with typed failure attribution, trigger explicit operations—Push, Pop, retry, or BacktrackTo the parent. We evaluate VBL with Qwen3.5‑9B, GPT‑4o, and DeepSeek‑V3.2 on AppWorld, ALFWorld, and ScienceWorld. VBL outperforms ReAct and reflection‑based control on every backbone‑benchmark pair, improving task success by up to 22.07 percentage points on the challenging AppWorld setting. Ablations show that removing either verification or diagnosis degrades performance, sometimes below ReAct, confirming the complementary nature of the two mechanisms. These results demonstrate that reliable long‑horizon interaction benefits from revising the planning state, not merely the next action.
Review: VBL’s structured goal stack and dual‑horizon checks provide a principled way to locate and recover from missing preconditions, offering a more robust control framework for long‑horizon language agents.