HackerNews中文版

我构建了一个名为 SafeAgent 的小型 Python 库，旨在保护 AI 代理在重试工具调用时避免产生现实世界的副作用。在实验代理工作流时，我们遇到的一个问题是，重试可能会多次触发不可逆的操作：代理调用工具 ↓ 网络超时 ↓ 代理重试 ↓ 副作用发生两次示例： • 重复支付 • 重复发送邮件 • 重复生成票据 • 重复交易大多数系统通过在不同服务中散布幂等性密钥来临时解决这个问题。 SafeAgent 将这一过程集中到一个小型执行保护机制中。这个想法很简单： 1. 每次工具执行都会获得一个 request_id 2. SafeAgent 记录执行凭证 3. 重试时返回原始凭证，而不是再次执行副作用示例：第一次调用真实副作用：发送邮件第二次调用（使用相同的 request_id） SafeAgent 返回原始执行凭证（没有第二次副作用）该项目仍处于早期阶段，但包括以下示例： • OpenAI 工具调用 • LangChain 风格的工具 • CrewAI 操作 PyPI: [https://pypi.org/project/safeagent-exec-guard/](https://pypi.org/project/safeagent-exec-guard/) GitHub: [https://github.com/azender1/SafeAgent](https://github.com/azender1/SafeAgent) 我很好奇其他人是如何处理代理副作用的重试安全问题的。

查看原文

I built a small Python library called SafeAgent that protects real-world side effects when AI agents retry tool calls.One issue we ran into while experimenting with agent workflows is that retries can trigger irreversible actions multiple times:agent calls tool ↓ network timeout ↓ agent retries ↓ side effect happens twiceExamples:• duplicate payment • duplicate email • duplicate ticket • duplicate tradeMost systems solve this ad-hoc using idempotency keys scattered around different services.SafeAgent centralizes this into a small execution guard.The idea is simple:1. every tool execution gets a request_id 2. SafeAgent records the execution receipt 3. retries return the original receipt instead of running the side effect againExample:FIRST CALL REAL SIDE EFFECT: sending emailSECOND CALL WITH SAME request_id SafeAgent returns the original execution receipt (no second side effect)The project is early but includes examples for:• OpenAI tool calls • LangChain style tools • CrewAI actionsPyPI: <a href="https://pypi.org/project/safeagent-exec-guard/" rel="nofollow">https://pypi.org/project/safeagent-exec-guard/</a>GitHub: <a href="https://github.com/azender1/SafeAgent" rel="nofollow">https://github.com/azender1/SafeAgent</a>Curious how other people are handling retry safety for agent side effects.

展示HN：SafeAgent – AI代理副作用的精确一次执行保护器