Agent 与自动化 4.0 · 优秀 2026-06-25 · 论文

Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection

反对让 LLM 直接生成不可靠爬虫代码,改为输出 typed JSON 采集配置:六类 collector 分类法模板与工具函数约束静态 Airflow DAG 执行规则质检与结构化反馈修正138 任务上验证 taxonomy;80 个独立核源任务上执行阶段零 LLM token平均 wall-clock 最低,用中等 one-shot 质量换可复用确定性可验证的定时采集路径工程读法:模型只产配置,执行与验收走确定性管线

打开原文回到归档

Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection

摘要(中文)

反对让 LLM 直接生成不可靠爬虫代码,改为输出 typed JSON 采集配置:六类 collector 分类法、模板与工具函数约束、静态 Airflow DAG 执行、规则质检与结构化反馈修正。138 任务上验证 taxonomy;80 个独立核源任务上执行阶段零 LLM token、平均 wall-clock 最低,用中等 one-shot 质量换可复用、确定性、可验证的定时采集路径。工程读法:模型只产配置,执行与验收走确定性管线。

Summary (English)

Instead of free-form scraper code, a constrained verifiable agent framework shifts LLM output to typed JSON collector configurations, combining a six-type collector taxonomy, template/utility constraints, static Airflow DAG execution, rule-based quality checking, and structured feedback correction. On 138 tasks the taxonomy supports description-based requirement typing; on 80 independently source-verified tasks the framework runs with zero execution-stage LLM tokens and lowest average wall-clock time, trading moderate one-shot quality for a reusable, deterministic, verifiable path for repeated scheduled open-web collection.

One-liner

网页采集 Agent:模型只写 typed JSON 配置,执行交给确定性 DAG 与规则检查。

Source body / metadata

arXiv abstract grounded intake for 2607.00035. PDF: https://arxiv.org/pdf/2607.00035v1

Instead of free-form scraper code, a constrained verifiable agent framework shifts LLM output to typed JSON collector configurations, combining a six-type collector taxonomy, template/utility constraints, static Airflow DAG execution, rule-based quality checking, and structured feedback correction. On 138 tasks the taxonomy supports description-based requirement typing; on 80 independently source-verified tasks the framework runs with zero execution-stage LLM tokens and lowest average wall-clock time, trading moderate one-shot quality for a reusable, deterministic, verifiable path for repeated scheduled open-web collection.

Obsidian evidence

  • local_note: OpenClaw定时任务/研究写作流水线/2026-07-22-研究写作流水线.md
  • intake_run: daily-intake-evening 2026-07-22