Execute–Verify–Commit: Action-level Verification for Long-Horizon Agents

In long-horizon tasks, a tool call can execute successfully while making a change that violates the user's instructions or the environment's operating rules. Such changes may compound across subsequent actions and leave the environment in a state that is difficult to recover from, causing the agent to spend its remaining steps on unsuccessful retries. We introduce Execute–Verify–Commit (EVC), a training-free method for action-level review before state changes are committed to the live environment. EVC executes a proposed state-changing action in an isolated fork and reviews the resulting tool output and state changes against user instructions and applicable policies, using the interaction history as context. Approved changes are committed without re-executing the action; rejected changes are discarded, and a specific explanation guides the agent's next step from the unchanged live state. With a fixed LLM reviewer and one review per candidate action, EVC improves average task success over no-review baselines across three top-tier acting LLMs-Qwen3.8-max, GLM-5.2, and Kimi-K3-on three benchmarks. Further experiments show benefits from execution evidence and specific rejection feedback, with little consistent gain from repeated voting. The discriminative reviewer Jev offers lower per-review cost but lower task success than the Qwen reviewer.