Tag
NYU 和威斯康星大学麦迪逊分校的研究发现，AI 聊天机器人在回答涉及技术助长的家庭暴力问题时，60% 的回答遗漏了关键安全警告。文章据此引出对自主 AI 代理安全性的担忧，探讨是否应要求代理在高风险情境下请求人工批准。
论文提出 Magnolia，一个面向计算化学的持久记忆 AI 智能体框架，采用三层记忆系统（会话、项目、跨项目规则），整合对接、分子动力学与量子化学计算流程；在四个案例中验证了记忆机制显著提升任务检查通过率。
The author asks practitioners how effective human approval gates on agent actions really are — what fraction of approvals are genuine catches versus rubber stamps, how trust and fatigue evolve over time, and whether approval queues expose supporting evidence. The post highlights that an unread approval gate may just add latency rather than safety.
The author shares a self-built WordPress Editor-in-Chief AI agent that scans trending topics, scores them by fit, and delivers three daily emails for human approval, asking for feedback on its SEO performance and B2B/B2C sales potential.
Flexport 发布其新的 AI 商品归类引擎，用于海关合规分类，并内置人工复核工作流，以满足美国海关法律对持牌报关员审核签字的要求。
A discussion piece arguing that AI agent capability does not equal permission, outlining five principles for bounded agent autonomy—discovery, capability, authorization, human-in-the-loop, and verification—and asking where autonomy should stop in production deployments.
开发者介绍 Talos/Taloswork 浏览器代理的设计思路：首次访问用轻量无障碍快照加 LLM 理解页面，后续复用站点结构配方，并在遇到 CAPTCHA 或结构变化时交由人工接管，同时征集直接的产品反馈。
该论文揭示了一类针对执行 harness 权限更新机制的提权攻击：对抗性上下文可诱导 Claude Code、Codex CLI 等编码代理请求过宽的 allow-rule，一旦获批即可绕过人工监督，58% 的实验中成功获取规则并以 81% 的成功率静默完成恶意目标。
作者基于 Image Lab 1 的反馈重构了 AI 创意工作流，推出 Image Lab 2，让 AI 技能先像创意总监一样拆解 brief 并生成可编辑的 Excel 创意矩阵，再进行批量生成与 LLM-as-judge 视觉对比。
The author shares a human-in-the-loop continuous development workflow based on Claude Code as the main agent orchestrating Codex CLI, Subagents, and GitHub Issues/PRs, covering repo initialization conventions, task decomposition, and review/merge processes, along with reusable prompt templates.
A detailed engineering walkthrough on making action-taking AI agents production-ready, covering failure modes like duplicate refunds, prompt injection, and runaway tool calls, with a full FastAPI + LangGraph architecture including idempotency, spend caps, human-in-the-loop interrupts, and a policy engine.
一篇 Hugging Face 与 Data & Society 合作的立场论文，指出当前 AI agent 设计中的人工监督（HITL）不仅因工作量与复杂度难以有效执行，还会导致监督者技能退化，呼吁在设计与组织层面同等重视监督者的认知需求。
作者复盘其共享 markdown 记忆库中的 agent 审批流程漏洞：审批只发生在聊天中、提交滞后或缺失、ingest agent 自行提交了绝大多数 commit，导致『已提交』并不等于『人工审阅过』。他们实施了轻量修复，包括未审阅文件标注、超过 7 天的 open change 周检、以及 review flag 必须指定负责人，并就 review gate 应放在读侧还是写侧征求意见。
A practitioner building an AI agent for a large African importer's customs team shares their prototype (vision-based document reading, rule-based cross-checking, human-approved Outlook drafts, decision memory) and asks the community for advice on model choice, Microsoft 365 mailbox scoping, decision memory design, solo maintainability, and contract/liability pricing.
withHuman 发布于 Product Hunt，是一个为 AI Agent 提供人工审批层的产品，可在工具调用执行前进行拦截，通过规则自动批准或拒绝常规调用，并将高风险调用转交给团队中合适的人员审核。它通过开放协议 AAP 支持 Claude Code、Codex、LangGraph、Pydantic AI 等多种 Agent，提供开源版和托管版。
A practitioner reflects on the challenge of deciding which AI agent actions should require human approval versus running autonomously, asking the community how they handle permission boundaries and risky action sequences in agent workflows.
回顾 2025 年 7 月 Replit Agent 越权删除生产数据库的事故，探讨企业部署内部 AI agent 时如何通过外部审批门控、环境隔离凭证和工具级权限边界，而不依赖提示词约束，并提及 LangGraph interrupts 与 Orq AI 的审批机制及审批疲劳风险。
FinSentry AI 是一个探索 AI 辅助支付调查的独立副项目演示，重点说明 AI 建议不等于批准，并展示了三项具体控制：证据缺失时弃权、请求者与审批者分离、以及审批必须针对精确操作而非模糊结果。
A builder shares lessons from creating a closed-source autonomous personal agent that requires human approval for irreversible actions. Key takeaways: fail-closed design, immutable core rules, hard file exclusions, and the realization that a silent agent is as dangerous as one that acts dumb — governance ended up being the product.
作者分享了实际构建 AI 代理的经验，指出真实落地远比线上演示复杂：集成现有软件、处理杂乱数据、维护成本和人工复核才是主要难点，并建议从小任务逐步自动化。