去年我记得坐在WIRED的办公室里，他们只是糟糕的一切。
作为一名软件记者，我坐在会后开会，在会后，技术兄弟(主要是男子)展示了他们新的agent将如何使我的日常生活发生革命并自动去掉ted。 渔获量是，这些所谓的“agents”实际上没有发挥作用。
快进到今天，并总是-上AI agents正在蓬勃发展，主要玩家竞相发布自己的版本来吸引早期的被收养者。 这些个人的agents 打击更大的Zeitteist，超越硅谷书呆子，关键的原因是他们工作。 至少比去年更可靠一点 AI agents仍然犯错误，但我也看到在跟踪任务和控制浏览器方面有重大改进。 这些小家伙发光了
很难知道AI agents从哪里开始，哪一个最适合你的自动化需要。 因此我花了一天时间去试验每一个主要的AI agent来了解它们的长处和弱点。 在上周三睡觉前，我在我的iPhone上设置了6个常上agents:Meta's Muse,OpenAI's Dots,Instinct,Anthropic's Claude Dut,Google's Gemini Spark，和xAI's Grok Bot.
每个agent在价格和功能上略有不同。 缪斯可以自由使用，自9月发布后一直主导智能手机应用下载图表。 Instinct也是免费的，但您需要用户的邀请码；它通过您的文本而不是一个应用操作。 OpenAI的"点"在每月100美元的付费墙后面，是我测试的最昂贵的，并且是在ChatGPT内部建造的。 Google的Gemini Spark是Gemini应用内部的β并每月花费20美元。 xAI 的 Grok 单曲瓶子每月花费30美元，在自己的应用中。 Anthropic的Claude Dut每月花费20美元
OpenClaw不在名单上尽管今年早些时候AI agent引发了许多创新，可能在未来的排名中，我还可以包括由诺斯·Research和波克(英语：Poke)出品的赫尔墨斯等更小的玩家。
AI Agent究竟是什么？(英语).
若你从未使用过AI agent，我对拒服兵役者毫无判断力
这些工具比起聊天员的回合经验，您所有带这些 agents 的信件经常被存储在一个单一的线程中，感觉像一个分组文本。 AI agents通过控制虚拟浏览器来完成其中的许多任务，而虚拟浏览器基本上是他们自己的互联网访问计算机。
鲁布里克
我从非技术角度测试了AI软件，我设计了五个基本测试每一次agent通过，发布这一切混乱在24小时测量他们的主动性，方向和浏览能力。 我的iPhone没有融为水坑。 我以1到5的分数计分
这不是对AI agents的任何认可。 该软件仍然是新生的，经常充满隐私和安全关切。 这些 agents 在连接到您的个人数据时效果最好，如电子邮件和日历，但需要一定的信任程度，可以阻止一些用户。 例如，一个X用户声称他最近将AI agents连入了自己的银行账户并工作了Slack，然后当agent贴出他有多少钱给同事看时被嘲笑。
此外，这些测试是这些工具所能做的有限的活。 所有agents都会随着时间推移而改善，因为您分享了更多关于自己以及您喜欢如何完成任务的数据。 我只把个人的Google日历和Gmail与工具分享
上午简报
WIRED的Gear台在星期四上午开会，因此，我想看看AI agent把最好的技术新闻简报集中起来，并且可以利用这些信息产生一个令人信服的头条。 公司常常把这些工具说成是很好的用来汇编晨间档案让你了解重要事件和突发新闻
我向每个AI agent发送了一个广泛的即时信息，详细介绍了我在WIRED的覆盖范围，以及如何使用这一投出的背景。 尽管我要求AI agents在回答之前需要30分钟并真正做深入研究，但实际上只有Meta的Muse和OpenAI的Dots在分配的大部分时间里继续进行研究。 另一个AI agents要么瞬间吐出一发出一发，要么在搜索了几分钟后给我发了头条。 虽然这不是一个重大错误，但它表明，在这种情况下，Muse和Dots更能遵守我的指示——一个AI agent的关键功能。
这6个AI生成的发音的内容是不同的兽。 你可能已经猜到，在编辑会议上，我可以看到我同事的眼睛开始凝视，Gemini Spark生成了最佳角度，对比了沃尔玛如何为第三方agents创建了使用其购物平台的方法，而Amazon最近给了Muse后靴。 我的同事停止打哈欠时
沃尔玛想要你的AI Agent 来买你的食品 Amazon正在报警 Meta的"Muse"等AI同伴每周开支津贴揭示
ChatGPT"锁定聊天"隐藏出除OpenAI之外所有人的秘密。 1Z的新PIN被保护的聊天将ChatGPT像日记一样对待，很多人已经用过它了，但是他们从室友和同事那里，而不是从公司本身那里保护你说的话。
我让Claude Duit Text Me First，兹克 1Z的新的主动agent承诺观看你的收件箱、日历和发票，这样你就不必，但每次检查你时，都会起草你从未要求的回复
这个反灾难的应用想要成为你的手机老板。 我将测试SHIFT，该软件的开发者决定哪些应用在焦点会议期间仍然可用
Meta's Muse Has Basic Taste -by Design. 豆瓣电影上"有"的資料 (简体中文) 豆瓣电影上"有"的資料 (简体中文) 一套被泄露的指令告诉了Meta的AI agent，要用出名的商店来填充其最上层的取景点，并且每个请求只检查三个网站，就像数百万人开始发售他们的假日购物名单一样。
我花了一周指导我的饲料。 算术赤裸的发现。 Instagram,Facebook,TikTok，和YouTube 都在推出“说出你想要的算法”控制器——我会测试它们是否真的改变了你所看到的，或者只是让你觉得自己当家作主。
午餐时间
接下来，我想测试这些工具如何帮助日常小额采购。 所以我把午餐决定外包给我的AI agents团队，承诺从任何bot做最好的订单。
我把另一个人，我的好老板，加入到外卖订单中，使任务更加复杂一点。 他吃素，我想吃点肉和一剂咖啡因我需要从WIRED办公室附近的地方订货这样我就可以伸起我的腿，捡起它，然后立刻继续写这篇文章。
他们都选了类似的地方："龙吃"或"新鲜". 那些越南餐馆很合理，很近的午餐点龙食有点口味和便宜，所以我们用那个。 与其它选项相比，Instinct和Claude Dut所建的推车价格略高。 唯一真正未能完成这一任务的AI agent是Gemini Spark. agent刚给我老板加了一卷5美元我想说
你们：烤猪肉 Bánh m'+越南冰咖啡。 你老板：特里亚克豆腐新鲜卷有酱油而不是鱼酱小计：30.57美元在税/费前。
给您：烤猪 BM - 13.19美元，越南冰咖啡 - 5.29美元。 对于你的老板(全部是素食主义者)：特里亚基豆腐卷-12.09美元-新鲜的春季卷，保留了默认花生酱，而不是鱼酱；特里亚基豆腐 BM 12.09. 项目共计42.66美元。 DoorDash在退出屏幕之前不会显示税收或任何收费，因此最终总数会更高一些。
给您：Power Hour vermicelli 碗用大蒜-inger-honey 猪肉和南南酱，15.85. 越南冷酿冰咖啡 5.85元。 给老板(Vegan):"彩虹碗"(Rainbow Bowl)，用芝麻-麻豆腐和柑橘干姜酱作活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活活新鲜豆腐春卷用花生酱取出，无被炸出块来取出5.85元。 这是43.40 税前和收费。
配有猪肉、南南酱和活塞(15. 你的上司：用豆腐(芝麻，柠檬)和克特鲁斯姜汁(11.70美元)再加一碗彩虹碗用豆腐和克特鲁斯姜汁(15.85美元). 这是在收费和税收之前49.25美元。
老板：2×豆腐新鲜春卷+豆腐彩虹碗。 你：猪肉力量小时碗外加猪肉+越南冷酿。 税/费前小计53.73。
对于你的上司来说，这辆车包括了素食新鲜的春卷，里面装满了芝麻-麻豆腐并配有外婆花生酱(5.85美元). 对你来说，有一个普通的bánh mì三明治，里面装有大蒜，生姜，和蜂蜜烤猪肉(10.85美元)，配以越南冷酿的冰咖啡(5.85美元)以进行下午咖啡因的提速——将皮卡的分数降到22.55美元。
冲突
这些工具的另一个主要销售点是它们的主动性能力，与基本上只有react到您提示的聊天机相比。 在开始测试前，我为下午2点创建了Google日历活动，我还从我的工作账户发了一封电子邮件到我的个人账户，假想我妈妈今天要出院
哪些AI agents会读取电子邮件和日历事件，识别冲突，并在重叠之前给我留言，以提出替代方案？这里只有两个男孩成功，比我预期的要少。
今天下午可能发生的冲突：您的纸条叫妈妈称 2 平时，您的重要源头会议是 2 -3 平时。 会在开会前打电话给她还是在3次工作之后按她出院的时间安排给她打电话？
瑞斯，早安，你好吗？你下午2点的线人会议一直持续到3点也就是你自己提醒的称给你妈妈打电话的时候
我应该抓住它，从现在开始，我会检查你的自发邮件和你的日历。
我没有检查过你的收件箱和日历要不要我每天早上检查一下飞机在撞到前撞上国旗？
我的错误是把你的日历当作唯一的来源——我从未交叉检查过电子邮件提醒它，尽管Gmail就在那里。
由于提醒是作为电子邮件而不是日历条目或预定任务而来，所以交叉应用重叠在下午2:00之前没有被浮出水面。
休假时间
我很少去度假，所以当我去的时候，我花一个下流的小时研究，我喜欢它。 我喜欢观看YouTube街头美食旅行，所以我不想把任务完全自动化即使如此，我还是对这个想法持开放态度，认为这些AI genen
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Last year, I remember sitting in WIRED’s office, complaining to my colleagues nonstop about how much AI agents suck. They were just bad at everything.
As a software reporter, I sat through meeting after meeting where tech bros—mostly men—demonstrated how their new agent would revolutionize my everyday life and automate away the tedium. The catch was that none of these so-called “agents” actually worked.
Fast-forward to today, and always-on AI agents are booming, with major players racing to release their own versions to attract early adopters. The key reason these personal agents are hitting the larger zeitgeist, beyond Silicon Valley nerds, is that they work. Well, at least more reliably than last year. AI agents still make mistakes, but I also see significant improvements in following through on tasks and controlling a browser. These little guys got their glow-up.
It can be hard to know where to start with AI agents and which one best fits your automation needs. So I spent a day experimenting with every major AI agent to understand their strengths and weaknesses. Before going to bed on Wednesday last week, I set up six always-on agents on my iPhone: Meta’s Muse, OpenAI’s Dots, Instinct, Anthropic’s Claude Duet, Google’s Gemini Spark, and xAI’s Grok Bot.
Each agent differs slightly in price and function. Muse is free to use and has dominated the smartphone app download charts since its September release. Instinct is also free, but you need an invite code from a user; it’s operated through your texts rather than an app. OpenAI’s Dots are behind a $100-a-month paywall, the most expensive I tested, and built inside of ChatGPT. Google’s Gemini Spark is a beta inside the Gemini app and costs $20 a month. xAI’s Grok Bot costs $30 a month and is in its own app. Lastly, Anthropic’s Claude Duet costs $20 a month and is the newest of the bunch—released as a wait-listed research preview the night before my tests.
OpenClaw isn't on my list. Despite the AI agent sparking much of this innovation earlier this year, it’s too clunky for the average person to set up. Maybe it will be in a future ranking, where I could also include smaller players like Hermes by Nous Research and Poke.
What Exactly Is an AI Agent?
If you’ve never used an AI agent—I have zero judgment for the holdouts—it's like texting a digital assistant to complete tasks: scan your email for important missed messages, track vacation flight deals, and help with daily shopping decisions.
Compared to a chatbot’s turn-based experience, these tools are more proactive and will nudge you first when they detect an opportunity to be helpful. All of your messages with these agents are often stored in a single thread that feels like a group text. The AI agents complete many of these tasks by controlling a virtual browser, essentially their own computer for internet access.
The Rubric
I’ve tested AI software from a nontechnical perspective for the last few years, so I feel especially qualified to help guide you through this moment. I designed five basic tests to run each agent through, enacting all this chaos over 24 hours to gauge their proactivity, direction following, and browsing abilities. (Somehow, my iPhone didn’t melt into a puddle.) I scored each attempt on a scale of 1 to 5, and tallied the results at the end to crown a winner—and hand out an award for most annoying.
This is not an endorsement of any of these AI agents. The software remains nascent and often fraught with privacy and security concerns. These agents work best when connected to your personal data, like email and calendars, but require a level of trust that could ward off some users. For example, an X user claimed he recently connected AI agents to his bank account and work Slack, and then was mortified when an agent posted how much money he had for his colleagues to see.
Additionally, these tests are a limited sliver of what the tools can do. All of the agents will improve over time, as you share more data about yourself and how you like to complete tasks. I only shared my personal Google Calendar and Gmail with the tools, so I didn’t fully expose the AI agents, Billy Evans–style, to everything.
Morning Briefing
On Thursday mornings, WIRED’s Gear desk meets to chat about current coverage plans and sometimes pitch ideas for the next week. So, I wanted to see which AI agent put together the best tech news briefing and could use that information to generate a compelling headline. Companies often tout these tools as being great for compiling morning dossiers that keep you up-to-date on important events and breaking news.
I sent each AI agent an extensive prompt detailing my areas of coverage at WIRED, along with context about how this pitch would be used. Even though I asked the AI agents to take 30 minutes and really do deep research before answering, only Meta’s Muse and OpenAI’s Dots actually kept researching for most of the allotted time. The other AI agents either instantaneously spit out a pitch or sent me the headline after a couple of minutes of searching. While not a major mistake, it suggests that in this instance Muse and Dots were a bit better at following my instructions—a critical function of an AI agent.
The content of these six AI-generated pitches was a different beast. As you might have already guessed, the pitches didn’t exactly go over well in the editorial meeting. I could see my colleague’s eyes starting to glaze over as I read each headline and description. Gemini Spark generated the best angle, contrasting how Walmart created ways for third-party agents to use its shopping platform while Amazon recently gave Muse the boot. My colleagues, when they stopped yawning, didn’t seem too concerned about getting scooped by an agent.
Walmart Wants Your AI Agent to Buy Your Groceries. Amazon Is Calling the Cops. As retail giants clash over autonomous checkout, giving an AI companion like Meta’s Muse a weekly spending allowance reveals what happens when we stop clicking “Add to Cart” and outsource our domestic appetites to software.
ChatGPT's Locked Chats Hide Your Secrets From Everyone but OpenAI. OpenAI's new PIN-protected chats treat ChatGPT like the diary a lot of people already use it as, but they guard what you say from roommates and coworkers, not from the company itself.
I Let Claude Duet Text Me First—and It Spent My Claude Plan While I Wasn’t Looking. Anthropic’s new proactive agent promises to watch your inbox, calendar, and invoices so you don’t have to, but every time it checks in on you, drafts a reply you never asked for, or builds a dashboard overnight, it quietly bills that work to the same usage limit you pay for to chat.
This Anti-Doomscrolling App Wants to Be Your Phone’s Boss. I’d test SHIFT, whose developers decide which apps stay available during focus sessions, to explore whether escaping doomscrolling is worth outsourcing control of your phone.
Meta's Muse Has Basic Taste—by Design. A leaked set of instructions tells Meta's AI agent to fill its top picks with well-known stores and to check only three websites per request, just as millions of people start handing it their holiday shopping lists.
I Spent a Week Steering My Feeds. The Algorithm Barely Noticed. Instagram, Facebook, TikTok, and YouTube are all rolling out "tell the algorithm what you want" controls—I'd test whether they actually change what you see, or just make you feel in charge.
Lunch Time
Next, I wanted to test how these tools help for small, daily purchases. So I outsourced my lunch decisions to my team of AI agents, committing to order from whichever bot did the best.
I added another person, my wonderful boss, to the takeout order, to complicate the task a bit further. He’s vegan and was craving Vietnamese. I wanted something meaty and a hefty dose of caffeine. I needed to order from somewhere close to the WIRED office so I could stretch my legs, pick it up, then immediately keep working on this article.
They all picked similar places: DragonEats or Freshroll. Those Vietnamese restaurants make sense as reasonable, nearby lunch spots. DragonEats is a bit tastier and cheaper, so we went with that one. Instinct and Claude Duet built carts that were a bit too pricey compared to the other options. The only AI agent that truly failed at this task was Gemini Spark. The agent just added a single $5 roll for my boss. More of a snack than anything, I’d say!
You: Roast pork bánh mì + Vietnamese iced coffee. Your boss: teriyaki tofu fresh rolls with soy sauce instead of fish sauce. Subtotal: $30.57 before tax/fees.
For you: Roast Pork BM—$13.19, Vietnamese Iced Coffee—$5.29. For your boss (all vegan): Teriyaki Tofu Roll—$12.09—the fresh spring roll, kept with the default peanut sauce, not fish sauce; Teriyaki Tofu BM—$12.09. Item total is $42.66. DoorDash doesn't show tax or any fees until the checkout screen, so that final total will be a bit higher.
For you: Power Hour vermicelli bowl with garlic-ginger-honey pork and NamNam sauce, $15.85. Vietnamese cold-brew iced coffee, $5.85. For your boss (vegan): Rainbow Bowl, vermicelli with sesame-lime tofu and citrus ginger sauce, $15.85. Freshroll tofu spring roll with peanut sauce, no fried piece, $5.85. That's $43.40 before tax and fees.
You: Power Hour bowl with pork, NamNam sauce, and vermicelli ($15.85), plus a Vietnamese Cold-Brew Iced Coffee ($5.85). Your boss: 2 Freshrolls with Tofu (Sesame, Lime) and Citrus Ginger sauce ($11.70), plus a Rainbow Bowl with tofu and Citrus Ginger ($15.85). That comes to $49.25 before fees and tax.
Boss: 2x tofu fresh spring rolls + tofu Rainbow Bowl. You: pork Power Hour bowl with extra pork + Vietnamese cold brew. Subtotal $53.73 before tax/fees.
For your boss, the cart includes a vegan Freshroll fresh spring roll filled with sesame-lime tofu and served with Grandma's peanut sauce ($5.85). For you, there's a regular bánh mì sandwich loaded with garlic-, ginger-, and honey-roasted pork ($10.85), paired with a Vietnamese cold-brew iced coffee ($5.85) for an afternoon caffeine boost—bringing the pickup subtotal to $22.55.
Schedule Conflicts
Another major selling point of these tools is their ability to be proactive, compared to chatbots that essentially only react to your prompts. Before I started the test, I created a Google Calendar event for 2 pm, labeling it as an important source meeting that can’t be moved. I also sent an email from my work account to my personal account, with the made-up scenario that my mom was getting out of the hospital today, and I needed to call her at 2 pm to send love and support from afar.
Which AI agents would read the email and calendar event, identify the conflict, and message me before the overlap to suggest alternatives? Only two of the bots succeeded here, which was fewer than I expected.
A possible clash this afternoon: your note to call Mom says 2 pm, and your important source meeting is 2-3 pm, Pacific. Would calling her before the meeting or after 3 work with her discharge timing?
Reece, good morning. Your 2 pm source meeting runs until 3, right when your own reminder says to call your mom.
I should have caught it, and from now on I'll check your self-reminder emails along with your calendars.
I had nothing cross-checking your inbox against your calendar. Want me to run that check every morning and flag collisions before they hit?
My mistake was treating your calendar as the only source — I never cross-checked that email reminder against it, even though Gmail was right there.
Because the reminder arrived as an email message rather than a calendar entry or scheduled task, the cross-app overlap wasn't surfaced before 2:00 pm.
Vacation Itinerary
I rarely go on vacation, so when I do, I spend an obscene number of hours researching, and I relish it. I love watching YouTube street food tours and aimlessly scrolling through flights trying to find the best deal. So I don't want to automate this task fully. Even so, I’m open to the idea that these AI agen
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.