Verified Tool Calls Improve LLM Agent Reliability Under Non-Atomic Failures

arXiv 2608.02645 反驳工具调用是原子的这一常见假设：现有 Agent 框架假设工具调用要么成功要么失败，但真实系统是超时的延迟可见的部分状态更新的一次发送出去没拿到响应，不代表没生效论文给的核心反直觉结论：纯重试基线在更高故障率下反而制造更多重复副作用；verify-only（只查不重试）的改进就比 retry-only 大；verify-before-retry 这个轻量包装器同时降低重复副作用并保住任务成功率论文在受控仿真环境里跨多个 task template 注入非原子故障做评估工程含义：错误响应操作失败，恢复决策必须基于外部状态而不是执行信号可观测外部状态幂等键兜底重试前先确认动作是否真没发生

Verified Tool Calls Improve LLM Agent Reliability Under Non-Atomic Failures

ID: 5eac797d
原文链接: https://arxiv.org/abs/2608.02645
作者/平台: Isham Kalappurackal Mansoor, Abhishek Phadke, Pratip Rana / arxiv
发布日期: 2026-07-31
归档分类: agents
标签: arxiv、agent-reliability、tool-call、verify-before-retry、idempotency
质量评分: 4/5
抓取时间: 2026-09-01T23:30+08:00

中文导读

为什么值得关注

工具调用不是原子的纯重试会制造重复副作用，verify-before-retry 才是正解

关键信息

文章标题：Verified Tool Calls Improve LLM Agent Reliability Under Non-Atomic Failures
作者/平台：Isham Kalappurackal Mansoor, Abhishek Phadke, Pratip Rana / arxiv
原文链接：https://arxiv.org/abs/2608.02645
发布日期：2026-07-31
关联标签：arxiv、agent-reliability、tool-call、verify-before-retry、idempotency

English Summary

arxiv 2608.02645 challenges the assumption that tool calls are atomic. Existing agent frameworks treat them as binary success/failure, but real systems exhibit timeouts after dispatch, delayed visibility, and partial state updates: a request may have taken effect even when no response came back. The paper shows that pure retry baselines produce more duplicate side-effects under higher failure rates; verify-only (just check, no retry) outperforms retry-only; and a lightweight verify-before-retry wrapper cuts duplicate side-effects while preserving task success. Evaluated in a controlled simulator with injected non-atomic failures across multiple task templates. Engineering implication: error responses are not operation failures recovery decisions must be based on external state (read-before-write diff), with idempotency keys as a safety net.

Obsidian Notes

来源：2026-09-01 AK-RSS Digest（89源精选）/ 每日综合摘要 / 调研 / DeepResearch 视所属主题而定
内容由 opencli 拉取原始来源 + Obsidian 笔记交叉核对生成。
中文导读与价值判断均锚定原文摘要与作者；未补充原文章节之外的细节。
抓取时间戳：2026-09-01T23:30+08:00。