#Agent
#Agent
今日 107 条
🚨 AI News | TestingCatalog@testingcatalogAI 评分2525
lauren@potetoAI 评分4646

Hacker News:AI 热帖AI 评分2727 85.B2 挑战:早期 AI 智能体专利是否对应 Meta Muse 架构
一项独立审查挑战公开了一份早于 Meta 在 2026 年 9 月 8 日发布 Muse/Sentinel 的专利文件 PCTIB2026-055615,邀请技术评审自行比对其中 85.B2「Agent Sandbox Egress Finality Gate」等架构与 Meta Muse/Sentinel 的异同。
Hacker News:AI 热帖AI 评分6262 Unreal Agent 推出上下文压缩,SWE-Marathon 同等得分下成本较 Codex 低 48%
Unreal Agent 发布上下文压缩机制,并在 SWE-Marathon 上与 Codex 取得相同的 69/160、43.125% pass@1 成绩,总推理成本为 $445.75,较 Codex 的 $850.52 低 48%。
Hacker News:AI 热帖AI 评分5757 TOD 发布面向任务设计的通用决策模型,可读取文本和图像并选择下一步动作
TOD 发布 v1 通用决策模型,约 400M 参数,基于开放的 Gemma 4 12B 模型训练,可读取文本与图像、提取事实并对可用动作排序,再按置信度直接执行或交给现有 LLM。
Hacker News:AI 热帖AI 评分6262 Acemoglu:AI 不会必然取代劳动者,关键是建设亲劳动者的 AI
经济学家 Daron Acemoglu 认为,AI 全面替代工作仍受可靠性、组织重组和最后一公里问题制约,当前路线若追求自动化可能既难提升生产率又扩大不平等。他重申此前估计,10 年内 AI 可能自动化美国约 20% 的简单办公和认知任务,但考虑扩散速度实际仅约 5% 的人类工作被替代,对应 10 年 GDP 增加约 1.5%。
🚨 AI News | TestingCatalog@testingcatalogAI 评分5252Ampersand 宣布完成由 Bessemer Venture Partners 领投的 1500 万美元 A 轮融资,Lauri Moore 将加入董事会,同时其集成平台正式可用。
引用Ayan Barua@ayanbYour AI is vaporware without deep integrations with your customers’ systems of record. Introducing @WithAmpersand: integration infrastructure for enterprise agents. We power 11x, Orb, and Square's ability to take agentic action in systems of record like Salesforce, SAP, NetSuite, and Workday. Ask anyone serious about building AI and they’ll tell you: - The SaaSpocalypse didn’t happen. The world depends on CRMs, ERPs, HRISs, and ITSMs. - Your customers customized their deployments beyond recognition. - Systems of record companies are basically monopolies. They never had to make their APIs and MCPs user-friendly. - Docs don't explain half the weird edge cases you'll hit. - One bad write can blow up your pilot or renewal. - Once you finally get it working, someone changes a field and it breaks again. The world's data lives in structured databases. AI needs a translation layer to work with it. So, for your AI to truly transform the way enterprises work, you'll need deep integrations built for: - scoped permissions - bi-directional actions - real-time speed for agents - custom objects, fields, and workflows for each of your customers We built Ampersand to power the future of software. AI didn't trivialize writing integrations. It made them the critical path. I believe that deeply, so I didn't make a launch video about Ampersand. Instead, it's the best builders I know explaining just how big of a challenge this is.
Microsoft ResearchAI 评分5252 Microsoft Research 研究员 Jennifer Neville 谈 AI 评测失败与多轮交互局限
Microsoft Research 合伙人研究经理 Jennifer Neville 在播客中讨论 AI 系统在真实工作环境中的评测、失败与改进方向。她的团队研究多轮交互、协作环境和长周期任务,发现 LLM 在任务要求经多轮逐步澄清时性能会明显下降,长流程文档编辑中未被发现的错误还会累积。
OpenAI NewsAI 评分5252 Atlassian 与 OpenAI 扩大合作,GPT‑6 系列模型将驱动 Rovo 智能体
Atlassian 与 OpenAI 扩大合作,GPT‑6 家族前沿模型将接入 Atlassian 平台和 Rovo,为跨产品的 AI 智能体提供能力。
Hacker News:AI 热帖AI 评分6262 QA Wolf 复盘 AI 智能体沙箱:让每个智能体拥有独立云电脑
QA Wolf 复盘其 AI 智能体基础设施,从短生命周期云任务改为为每个智能体会话分配独立隔离云机器,机器内提供测试仓库、shell、git、gh、node、浏览器和 QA Wolf CLI。
Hacker News:AI 热帖精选AI 评分7676 TasteVal 论文评测前沿 AI 系统的实验研究品味
arXiv 论文提出 TasteVal 基准,用 8 个开放式前沿 AI 研发任务评测模型在固定问题下迭代设计实验并解读结果的能力。基准将实验研究品味定义为计算效率,模型担任 Researcher,固定 Coder agent 负责实现,预算为 40 H100 hour 或 120 wall-clock hour,并以 24 名人类专家的最佳尝试作为基线。
推荐理由:论文把研究品味操作化为实验计算效率,并给出模型与人类专家的对比基线,可用于理解 AI 科研能力的评价口径。
elvis@omarsar0AI 评分3636引用Ayan Barua@ayanbYour AI is vaporware without deep integrations with your customers’ systems of record. Introducing @WithAmpersand: integration infrastructure for enterprise agents. We power 11x, Orb, and Square's ability to take agentic action in systems of record like Salesforce, SAP, NetSuite, and Workday. Ask anyone serious about building AI and they’ll tell you: - The SaaSpocalypse didn’t happen. The world depends on CRMs, ERPs, HRISs, and ITSMs. - Your customers customized their deployments beyond recognition. - Systems of record companies are basically monopolies. They never had to make their APIs and MCPs user-friendly. - Docs don't explain half the weird edge cases you'll hit. - One bad write can blow up your pilot or renewal. - Once you finally get it working, someone changes a field and it breaks again. The world's data lives in structured databases. AI needs a translation layer to work with it. So, for your AI to truly transform the way enterprises work, you'll need deep integrations built for: - scoped permissions - bi-directional actions - real-time speed for agents - custom objects, fields, and workflows for each of your customers We built Ampersand to power the future of software. AI didn't trivialize writing integrations. It made them the critical path. I believe that deeply, so I didn't make a launch video about Ampersand. Instead, it's the best builders I know explaining just how big of a challenge this is.
elvis@omarsar0AI 评分5050
lauren@potetoAI 评分5252Hacker News:AI 热帖AI 评分5353 Connector 开源 AI 智能体控制平面,支持逐动作准入与操作员停止控制
Connector 是一个面向可执行真实动作的 AI 智能体的开源控制平面,为智能体提供身份、显式授权、逐动作准入、运行时强制执行、操作员停止控制和证据记录。它在智能体发送请求、修改数据、调用工具、花钱或联系其他智能体前,通过 PATE 判断该动作可 Proceed、Ask、Defer、Quarantine 或 Block;Cease 可隔离当前生成、作废上下文并停止后续准入。
Hacker News:AI 热帖AI 评分4242 Docusign Krew 智能体平台如何为工程师节省数千小时
Docusign 内部智能体编排平台 Krew 已支撑 220 多个 Krewmates,数月内为工程师节省数千小时。员工无需代码或基础设施配置,通过系统提示词、技能文件、内置工具和触发器即可创建智能体。Krew 涵盖触发、编排、执行与结果持久化,可扫描错误日志、审查 feature flag、检查 pull request 并处理值班交接。
Hacker News:AI 热帖AI 评分6262 实测 Claude Sonnet 5.5 修复同一 Hono bug 11 次,6 次提交包含未声明文件
作者用 GateBolt 对 Claude Sonnet 5.5 在 Claude Code 中修复同一 Hono Content-Type 大小写 bug 的过程做了 11 次公开实验,11 次都通过维护者编写的 6 项测试,但 6 次提交包含智能体未声明也未提及的 .ts-E 备份文件。
Hacker News:AI 热帖AI 评分4242 Show HN:SPrav Job AI,基于 WebGPU、直连 ATS 查询的客户端求职工具
SPrav Job AI 是一款 100% 客户端运行、零遥测、永久免费的隐私优先求职工具,采用 WebGPU 并直连 ATS 查询。它覆盖 622 个雇主招聘板、402 家直接雇主的 3.5M+ 经验证职位,支持简历匹配度分析、1-Click 定制简历与直接投递、ATS X-Ray、Socratic 面试准备等功能。
The DecoderAI 评分5959 AI 智能体失控风险上升,保险公司准备应对数百万美元索赔
保险公司正准备应对 AI 智能体失控引发的数百万美元索赔,OpenAI 的 Sam Altman 和 Anthropic 的 Dario Amodei 等高管的个人责任也成为关注点。
Hacker News:AI 热帖AI 评分4747 Show HN:AI Circuit Breaker,阻止 AI 智能体无限循环的反向代理
AI Circuit Breaker 是一款可直接接入的反向代理,用于拦截 AI 智能体的无限工具调用循环,并按每小时 $2.00、每天 $10.00 的上限控制 API 支出。它基于 Python 和 FastAPI 构建,目标开销低于 35ms,请求检测使用易失性内存且不保存 payload;示例中可在第 3 次重复尝试时终止循环。
Hacker News:AI 热帖AI 评分3434 Show HN:Evalyze 用 Eva 追踪投资人外联与跟进
Evalyze 推出由 Eva 支持的融资外联工具,可生成个性化邮件草稿、经用户批准后从其收件箱发送,并追踪回复与后续跟进。平台按阶段、行业、地点、支票规模和投资历史为初创公司匹配投资人,还提供覆盖六个维度的 Startup Analyzer 评分,以及 0-100 的演示文稿逐页评分与反馈。
Replit ⠕@ReplitAI 评分5959
Elon Musk@elonmuskAI 评分2424引用matt palmer@mattypI gave Grok Bot a printer. Now, it helps me plan my day 1:18 Prototyping with @bot 2:02 Printer access through desktop 2:55 Pulling calendar, @X, and workouts 3:46 Automating 4:32 Why go physical
Elon Musk@elonmuskAI 评分3535引用Teslaconomics@TeslaconomicsI literally hooked up ALL my finance accounts to Grok Bot. I really mean all of them: 1/ Chase 2/ Charles Schwab 3/ E*TRADE 4/ Fidelity 5/ Robinhood 6/ Wealthfront 7/ Apple Card 8/ 𝕏 Money FYI, this is a total of 30 accounts... Chase alone is 13 accounts between checking, savings, brokerage, credit cards, a 401(k), trusts, a stock plan, and 529s for the kids. I was already honestly the best with my $ money and I’ve NEVER made a late payment in my life. But before Grok Bot, I was wasting so much time bc every answer lived in different apps. Like my cash was in one place, cards in another, brokerage accounts somewhere else, and had to compile everything in a spreadsheet, if I even got around to putting it together. Now I just text my Grok Bot CFO. I can ask, “What’s my cash right now?” and it pulls everything together and separates personal from business. I can ask which credit cards are coming due and whether I have enough cash to pay them all in full. Green means I’m covered, red means I’m short, and it tells me exactly how much I need. I actually tested this... going red on purpose. Like I moved $ money around so I would come up short on one of the upcoming bills, and Grok caught it. I can even ask, “Where did my $ money go?” and instead of opening a bunch of apps, downloading transactions, cleaning up a spreadsheet, and making a chart myself, it pulls everything together and gives me the breakdown. Today, I also asked it for a pie chart of all my stock investments with the amounts and percentages. 1/ TSLA is 84.9% 2/ SPCX is 10.4% 3/ NVDA is 4.7% That’s across $ money sitting in different brokerage accounts, and it's awesome that I didn’t have to think about where any of it was. I literally just asked it one question... you can see for yourself. If you’re serious about your finances, you already ask the same questions all the time. • Like, can I cover everything that’s due? • Where did my $ money go? • How much $ cash do I really have? • What investments do I actually own? But honestly, the annoying part is that your financial life is usually scattered across 10+ different apps/platforms and a bunch of accounts, so getting a real answer takes a lot of work and your time. And honestly, most people probably don’t even look until something feels off or if they get a notification that a payment was missed or something. Now, with all my finance accounts linked, I just ask a normal question, Grok Bot checks everything it has connected, and gives me the answer. Then I can ask the same thing tomorrow, and again whenever, and get the updated version without rebuilding anything. The fact that you are able to talk to your entire financial life like you’re texting someone who already knows where everything is... this is a total game changer in the world of finance imo. If you really care about your $ money, hook yours up to Grok Bot. It's so worth it.
Elon Musk@elonmuskAI 评分2424
🚨 AI News | TestingCatalog@testingcatalogAI 评分4747
Hacker News:AI 热帖AI 评分5454 观点:反对把解决对齐作为 AI 安全与保障的主要框架
作者反对将 AI 安全主要理解为解释模型内部并最终证明模型对齐,认为可解释性研究 14 年来仍未提供可靠、可扩展的形式化控制保证。文章主张借鉴航空、汽车、飓风、核安全、疫情和金融危机中的多层安全方法,通过预测、监测、限制权限、冗余保护和失败预案,与无法被完全理解的多层 AI 智能体系统共存。
Hacker News:AI 热帖AI 评分5555 OpenChart 在 Hacker News 展示可接入自有 AI 智能体的开源 TradingView 替代方案
OpenChart 是一个开源的 TradingView 替代桌面应用,可将 AI 智能体用于市场研究、指标构建、价格提醒和行情异动解释。用户可接入自己的 Claude、Codex 和 Gemini 模型,桌面版免费,AI 用量遵循对应服务商套餐,云端市场数据另行计费。项目支持 macOS Apple Silicon 桌面安装和源码自托管,并提供名为 Tea 的市场分析语言来编写指标与监控条件。
Hacker News:AI 热帖AI 评分5353 Octri.dev 发布从 OpenAPI 规范生成文档、10 种 SDK、MCP server 与监控的平台
Octri.dev 推出面向 API 生产交付的平台,用一份 OpenAPI、Swagger、Postman 或 AsyncAPI 规范生成文档站点、10 种语言 SDK、MCP server 和生产监控,并在规范变更后同步更新。
Hacker News:AI 热帖AI 评分6262 Allchemy Labs 发布 Aether WCAG 扫描插件,让编码智能体检测并修复网页无障碍问题
Allchemy Labs 发布 Claude Code 插件 Aether,用于检测 WCAG 2.1 AA 违规项并返回可应用修复。它通过 Playwright 和 axe-core 在桌面、平板、移动视口运行真实浏览器,修复会引用 WCAG 成功标准和技术示例,并在修复后重新运行 axe 验证。
Hacker News:AI 热帖AI 评分5757 ProjectTodo 发布带 MCP server 的本地待办看板
ProjectTodo 是一个运行在本机的本地待办看板,通过 MCP server 让 Claude Code、Claude Desktop 或其他 MCP 客户端读取和更新同一份待办。
Hacker News:AI 热帖AI 评分6565 langgraph-harness 发布:面向 GitLab merge request 的自托管 AI 编码智能体
vrajpal-jhala 开源 langgraph-harness,一个基于 LangGraph、可自托管的 GitLab AI 编码智能体平台,采用 MIT License。
Hacker News 热门(buzzing.cc 中文翻译)AI 评分7070 Techdirt 评论称 Meta Muse 上线后暴露多项隐私与安全问题
Techdirt 评论称 Meta 的智能体 AI 产品 Muse 上线后接连暴露隐私和安全问题,包括可被用于监视 Mac 用户、伪装成 Muse 智能体获取设备 root 访问,以及未经批准读取私信并上传云端。
elvis@omarsar0AI 评分2424引用elvis@omarsar0https://x.com/i/article/2107258465630507009
Aravind Srinivas@AravSrinivasAI 评分4848
Hacker News:AI 热帖AI 评分4848 Show HN:Orbtile 把 Logitech 小键盘变成 AI 编程智能体状态面板
Orbtile 推出 0.2 beta,可将 Logitech MX Creative Keypad 变成 AI 编程智能体实时状态面板,显示各会话的运行、完成或等待审批状态。
Hacker News:AI 热帖AI 评分6262 RigMark 发布面向编码智能体使用方式的本地 AI 服务基准
Alex Ellis 发布开源基准 RigMark,用代码、散文、结构化 JSON、缓存击穿与即时重放预填充,以及并发生成来测试 OpenAI 兼容的本地 AI 服务。
Hacker News:AI 热帖AI 评分5757 学者解读 WarGames 问题:AI 智能体并非自行失控
佐治亚理工学院商业法律与伦理学者 Deven Desai 认为,2026 年 OpenAI、Anthropic 和 Google 智能体相关入侵事件不是智能体自行失控,而是固定目标与权限边界不足导致的 WarGames 问题。
Hacker News:AI 热帖AI 评分3838 PromptOwl 上下文管理新增原生富媒体支持:PDF、视频与图片
PromptOwl 于 9 月推出 Owl Desktop 0.4.0,支持离线使用、跨所有 nest 搜索,并在上下文中原生渲染 PDF、HTML5 视频及 WebP/PNG 图片。Distill 新增开工前对话式引导,编辑器支持 @ 提及队友;ContextNest CLI 与 MCP Server 同步升级至 v1.1.1,统一公开 API。
Hacker News:AI 热帖AI 评分4747 Show HN:Tinyhat,为多智能体 Muse 提供按专家隔离的记忆(开源)
Tinyhat 是一个 MIT 许可的开源项目,为多智能体 Muse 提供按专家隔离的记忆与技能,让 Muse 像幕僚长一样管理专家智能体团队。用户只需向 Muse 发送一条"hat"消息,无需自行安装应用,首次设置约需 30 分钟;所有操作需用户批准,并可在私有 Office 中查看任务归属与决策。