中国版Muse能管事！我把宠物保单、双11算价和周末旅行全交给了灵光2.0
最近两个月，Personal Agent 这个概念忽然从边缘话题变成了普通用户手机屏幕上的新图标。
Meta 在 9 月 8 日正式推出 Muse，短短 40 天内下载量突破 1000 万次。随后，OpenAI 推出 Dots，Anthropic 推出 Claude Duet，Personal Agent 满地跑。
对于国内普通用户来讲，用上上面的 Personal Agent 要折腾半天，加上刷了一些翻车案例，把个人生活的各种权限交给 AI 让我有点顾虑。
关于 Personal Agent ，我在意的是，离开对话框后，能不能放心让它帮我管事。
今天，蚂蚁正式发布了灵光 2.0，APPSO 体验下来，感觉它最不一样的地方就是「能管事」。
灵光 2.0 将 AI 介入任务的起点进一步前移，能够主动识别潜在需求、持续跟进复杂事务，并在关键节点及时提醒。
事情发生变化时，它会自己发现、及时调整并推进新的规划，全程主动、自动。也正因为如此，它最接得住的是一类过去 AI 很难办好的活——时间跨度长的长程任务、信息一直在变的任务、需要反复记住前情的任务。
举个例子🌰：当我对它说「下周去上海出差三天」，订机票、订酒店、安排日程等只是起点；预订成功后，它会主动提醒我「该出发去机场了」、「别忘了带身份证」；如果途中遇到航班延误，它还会自动为我重新调整后续的酒店入住和会议行程。
作为一个纯过日子人，「维持生活」本身就是一项庞大的工程。而用了灵光 2.0 之后，日常不再只是被动承受，普通人也能在琐碎的柴米油盐里，打出自己的「战术」与创造。
这篇文章想记录的，就是我把哪些烦心事交给了它，以及它是怎么管到底的。
顺便说一下，灵光 2.0 已全面开放，大家可以前去体验。
家里的事：信息太多，靠记性是真的扛不住
我养了一只狗，几年下来，对信息密度这件事有了新的体会。打疫苗要看时间节点，买驱虫药要对照体重，生病了要先搞清楚保险的报销边界，省得白跑一趟医院，也省得账单出来才发现这项不在范围内。
这次我把宠物医疗险的合同发给灵光，顺带说了一下狗最近的状态：这两天一直不太吃东西，精神有点蔫，不知道是不是肠胃的问题。
我让它告诉我这种情况可能是什么原因，按合同规定，相关治疗能不能报销，该去哪个级别的医院，就诊时要留存哪些材料。
读完合同后，它逐条列出了对应的报销范围和凭证类型。它还补充了一些判断依据，告诉我如果情况没有好转，哪些症状是该去医院的信号，哪些观察一天也无妨。我顺手让它帮我设了一个跟进提醒，到时候再来问问情况有没有改善。
等它来提醒我时，我和它更新了后续，它展现出了极强的长期记忆能力，完整保留了狗狗之前的病况、历史进展和我的偏好。
我完全不需要每次重头交代背景，它直接基于先前的对话给出下一步建议。
类似的场景还有群消息。业主群、家长群、活动群，每天有大量消息滑过，需要我回应的，可能只有两三条。靠自己逐条盯着，实在费心力。让灵光接入钉钉，筛出需要跟进的内容，能省下不少注意力。
用下来我还发现一个规律，它拿到的上下文越丰富，智能体现得越强——接入群后，哪些消息需要优先回应，它分得很清楚。
钱的事：双 11 快来了，我让它帮我蹲低价
每年到双 11，我都会后悔。
要么是手快，提前几天下单，结果活动正式开启那天，价格更便宜；要么是手慢，心想再等等，打折窗口一过，价格又涨了回去。
不是没有心思去盯，是真的盯不过来——平台规则每年都在变，「预售价」「尾款日」「跨店满减」「店铺券」一层叠一层，要算清楚真正的到手价，得花上一番心思；更何况还要同时跟几件商品。
今年我换了个方式。我把最近想囤的几样东西整理了一下，截图发给灵光，让它帮我标出哪几件性价比最高，把折扣和满减都折算进去，给出最终到手价的排名。
它整理完之后，我追加了一个条件：预算控制在 500 以内，帮我重新排一次优先级。
当天晚上，我又追加了一个要求，让它对比一下不同平台的同款价格，看看哪边划算。然后让它持续盯着价格，有明显波动就提醒我，活动节点到了也提前说一声。
之后的事，就不需要我操心了。
第二天它主动发来了一条价格变动提示，说明了某件商品便宜了多少，比前一天低了多少，现在值不值得剁手。
对于一些流程比较长的任务，灵光 2.0 通过独立云主机自带电脑和浏览器执行得不错，即便我关掉 App，它也在后台继续跑任务、盯状态。
和我不一样，灵光不觉得重复无聊，不会因为价格迟迟不动就失去耐心，也不会在窗口期没注意到就让我错过。
这让我想到 Stripe 的 CEO Patrick Collison 曾提出的一个设想：Personal Agent 会让普通人第一次拥有过去只有企业才负担得起的信息处理能力，可以持续找优惠、逐条查条款、认真比商品。用户的注意力不再是被消耗的资源，好产品也更容易被认真对待它的人看见。
此外，用户可以授权接入支付宝，灵光不仅能梳理账单明细，还能精准识别容易被忽视的自动续费服务、剖析消费习惯、给出实用的省钱建议。
我让它帮我扫一遍账单找找周期性扣款，这才发现自己开了三个网盘会员，两个常年吃灰。
我立马停掉了自动续费，还让灵光帮忙分析了我的消费习惯，再决定哪些钱要花。
双十一的购物车已经按优先级排好 —— 这次该买的买，不该花的，一分不花。
如果你也好奇自己的钱都去哪儿了，不妨也让灵光帮你扫一眼，说不定你比我藏的「惊喜」还多。
出行的事：计划跟得上变化
如果你做过旅行计划，你最大的感触可能是：计划赶不上变化。
你花了一个晚上查攻略、排时间、订餐厅，方案刚做好，朋友说临时有事来不了，或者天气突然变脸，整个行程就得推倒重来。最烦的是我知道自己要重新规划，却总是一拖再拖——反正又要从头来过，就先放着吧。
我最近试着把这件事交给灵光。
灵光展现出的的跟进与调整能力还挺让我惊喜，遇到天气变化或行程变动，它会主动弹窗提醒，并基于新情况动态重新规划，自动调整原计划中的住宿、路线和节奏，把变化带来的麻烦降到最低。
我给它的原始需求是：这个周末，我和闺蜜，还有她家一个 4 岁的小孩想出去玩，开车两小时以内，行程不能太赶，孩子要午睡，我们都不能吃辣，周日晚上 8 点前必须到家，人均预算在 400 左右，帮我规划。
它出方案的过程中问了几个关键问题，然后给出了完整的行程安排：去哪里、上午做什么、午饭在哪吃、哪里方便孩子午睡、下午继续去哪、几点出发最合适。它还告诉我，周六出发前需要带什么，天气如果有变化会提前通知我。
第二天，我告诉它：闺蜜来不了了，就我一个人去，行程帮我重排一下，预算不变，但一个人可以更灵活，想多加一个地方。
它重新规划了，把原来的框架保留，把涉及两个人或小孩的安排调整掉，把新增的目的地填进去。一个人行动，不用预留午睡时段，节奏可以自由一些，可以去稍远一点的地方，它把这些因素都考虑进去了。
对话里积累的上下文它一直记得，你告诉它什么变了，它去更新相关的部分，不让你一直重复自己。
不过计划仍然赶不上变化，但以后变化来了，我大概不会再拖着迟迟不动手了——因为重新规划这件事，不再完全是我的事。
Muse 替你跑腿，灵光 2.0 帮你管事
社会会学者 Allison Daminger 在 2019 年的一项家庭分工研究中，把那些不容易被看见的「认知劳动」拆成四个环节：预判需求、寻找可选方案、作出决定、跟进结果。它们和做饭、打扫一样，维持着一个家的运转。
从这个视角来看 Personal Agent，会发现一个被忽略的前提：当我们给 AI 一条足够清楚的指令时，往往已经替它完成了相当一部分工作。
在 APPSO 看来，「管事」应该成为评价个人 Agent 的重要维度。除了完成指令，还要看它能不能分担指令前后那些识别问题、持续留意、接续安排的劳动。
几乎所有的 AI 都在卷「怎么帮用户干活」——写文章、画图、搜攻略、生成小应用。但这类交互本质上依然停留在「单次输入、单次输出」的工具层面。
灵光 2.0 开始尝试推动 AI 助理从被动响应指令、完成单次任务，向主动发现需求、持续管理事务演进。
用了几天，我发现不必再靠记忆和自觉去盯着每一件事。
Personal Agent 的价值，就体现在点击「确认」之前那些漫长、零碎的时间里。有 Personal Agent 核对信息、记住前情、及时跟进，我就不用一边忙着手上的事，一边担心还有什么没顾上。
你只需要说一次想要什么，剩下的搜索、比较、记录和提醒，它在后台慢慢推进。有消息了，它来告诉你；到了该你拿主意的时候，它把整理好的选项摆在你面前。
回看之前的「闪应用」，灵光已经在试着让普通人用自然语言，生成一个能交互、能修改的小工具，去满足每个人的具体而微小的需求。
以个人需求出发，让更多复杂生活琐事变简单的产品逻辑，也一直延续到这次的灵光 2.0。从这个角度看，它也是最早一批 Personal Agent 产品。
灵光 2.0 或许会让你想到同样主打个人生活场景的 Muse，但它们给我的感觉很不一样。如果说 Muse 是「跑腿型」Personal Agent ，帮我处理邮件、下单商品。那灵光 2.0 则是「管事型」Personal Agent，交给它的更多是生活里那些没完没了的小事。
这也呼应了今年「龙虾」热潮的教训：不少人部署之后找不到真实的使用场景，热度很快消散。
目前市面上的 Personal Agent 形态趋同，差异主要在入口和生态。
云端虚拟机、24 小时不间断运行，以及对敏感操作的二次确认，如今已成为 AI Agent 的行业标配。例如 Spark 在付费或发送邮件前会先征求许可，Dots 则将修改密码等高危任务交给用户亲力亲为。
在这套标准架构之下，玩家之间的核心差异在于各自手中握有的底牌：Google 靠的是邮箱生态，Meta 掌握着社交分发，OpenAI 拥有庞大的插件生态……
Personal Agent 能否打通服务、获得授权，并且安全可靠地完成执行，决定着 Agent 能否足够个人、个性化。
Muse 已接入 Stripe 的 Link 支付，并首个获得其购物保障，说明支付基础设施会是关键一环。
恰巧，蚂蚁灵光 2.0 所在的生态，与支付、生活服务等真实交易场景有天然联系，并且已经将这些优势转化为 Agent 的执行能力。
经过测试，灵光 2.0 不仅内置打通了支付宝（账单、信用卡、花呗、借呗、出行、会员、出行、车生活等）、QQ 邮箱、训记、日历、闹钟、短信、文件等应用服务与系统服务，还支持在开放环境中调用 MCP、CLI 等工具，跨平台可连接钉钉、飞书、滴滴、百度网盘、米家、理想/蔚来/小鹏/华为汽车、12306 及智能硬件设备等等。
Personal Agent 不应该让我们变得更忙，它得把我们从维持生活的忙碌里面解放出来。人当然应该保留最终决定权，也应该能放下随时待命的负担。
日子已经够费心了，那些悬而未决的小事，能有灵光这样的助理记挂，我也就能匀出一点心思，留给自己想过的生活。
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。
The Chinese version of Muse is in charge! I gave the pet policy, the double 11 and the weekend trip to Viola 2.0.
In the last two months, the concept of PERSONAL Agent suddenly turned from a marginal topic to a new icon on a regular user's mobile screen.
Meta officially launched Muse on September 8th, with over 10 million downloads in 40 days. Then Open AI launched Dots, Anthropic launched Claude Duet, Personal Agent ran all over the place.
For ordinary users in the country, it gives me some concern to use the personal Agent on the top for half a day, plus some cases of car overturning and handing over to AI all the permissions of personal life.
With regard to personal Agent, what I'm concerned about is if, after leaving the dialogue box, I can count on it to take charge of my business.
Today, the ants are officially publish in the light of 2.0 and APPSO experience, and it feels like the most different part is "manager."
Psychic 2.0 Moves further the starting point for AI intervention tasks, enabling proactive identification of potential needs, continuous follow-up of complex matters and timely reminders at key nodes.
When things change, it discovers, adapts and advances new plans in a timely manner, and is proactive and automatic. And that's why it's got a kind of hard-to-do activity in the past — long-distance missions over time, missions where information is constantly changing and tasks that need to be remembered repeatedly.
By way of example: when I say to it, "Shanghai travels for three days next week", the booking of a ticket, hotel booking, scheduling, etc. is only the starting point; when the booking is successful, it will voluntarily remind me that "it's time to go to the airport" and "don't forget to bring an identity card"; if there is a delay in the flight, it will automatically reschedule the stay and meeting schedule of my subsequent hotel.
As a mere living person, "support" is in itself a huge project. After using the light 2.0, everyday life is no longer just passive, and ordinary people are able to play their own "tactical" and "creation" in simplistic woody salt.
What this article is trying to record is what I've given it to worry and how it works.
By the way, the light 2.0 is fully open and you can all go for it.
Home stuff: Too much information, memory can't handle it.
I've had a dog, and for years, I've had a new experience of information density. Vaccination depends on the point of time, the amount of anti-worming medicine needed to be measured against weight, the amount of insurance coverage needed to be determined in the event of illness, the amount of money spent on a hospital, and the amount of money spent on it.
This time, I gave the pet medical insurance contract to the light, and by the way, the dog's state: I haven't eaten much for two days, and I'm a little upset and I don't know if it's a gastrointestinal problem.
I asked him to tell me what the reason might be, whether the treatment was reimbursable under the terms of the contract, which level of hospital should be used, and what materials should be retained at the time of the visit.
After reading the contract, it lists the corresponding reimbursement scope and type of certificate by article. It also adds a number of grounds for judgement, and tells me which symptoms are signs of access to the hospital if the situation does not improve, and which days to observe. I'll let it set me up a follow-up reminder, and then I'll ask if it's any better.
And when it came to remind me, I updated it, and it showed the power of long-term memory, and it preserved the condition of the dog, the progress of history and my preferences.
I do not need to return to the background every time, and it gives next step recommendations based directly on previous dialogues.
There's a lot of news on the scene. There is a lot of slippage in the ownership, parents and activities that may require my response. It's hard to keep an eye on each of them. Let the light access the nails, sift out what needs to be followed, and save a lot of attention.
I also found a pattern, the richer it gets, the stronger it gets context, the stronger it gets AI agent — the priority information that needs to be answered, and it's clear.
Money: Two, eleven. Come on, I'll put it down.
Every year to double 11, I regret it.
Either the hand was fast, the bill was placed a few days in advance, and the result was that the day the event was officially opened, the price was cheaper; or the hand was slow, the mind was waiting, the window was broken and the price went back up.
The platform rules are changing every year, and the "end-of-the-sale" "end-of-the-sale" "end-of-sale" "cross-the-street" "street-of-the-street" is one layer at a time, and it's going to take a while to figure out what the real-to-the-hand price is, let's add a few items.
I changed the way this year. I've sorted out a few things I've been trying to hoard recently, and I've sent a screenshot to the light so that it can help me to mark which items have the highest value for money, to convert both the discounts and the full amount, and to rank the final price.
When it was finished, I added the condition that the budget be within 500, and help me re-prioritize.
That night, I added a request to compare the same price for the different platforms and see which side to pay. Then let it keep staring at the price, and let me know if there's any obvious fluctuations.
After that, I don't need to worry.
The following day, it sent out an offer of price movements indicating how cheap a particular commodity was, how low it was the previous day, which was not worth cutting off.
For some of the longer tasks, the light 2.0 is well carried out by the independent cloud host with its own computer and browser, and even if I turn it off, it continues to run and stare backstage.
Unlike me, it's not boring to repeat it, not to lose patience because of the delay in the price, or to let me miss it without notice during the window.
This reminds me of the idea put forward by the CEO Patrick Collison of Stripe: that for the first time, Personal Agent would give ordinary people the ability to process information that used to be affordable only for businesses, to continuously look for concessions, article by article, and carefully compare goods. The attention of users is no longer consumed resources, and good products are more easily visible to those who take them seriously.
In addition, users can authorize access to the money to pay, not only to clean up the bills, but also to identify easily neglected automatic renewal services, to analyze consumption habits, and to give practical savings advice.
I asked him to clean up my bills and find a periodic deduction, so I found out that I had three network members, two of them eating ash.
I stopped the automatic revolving fee right now, and I asked Shinzi to help me analyze my consumption habits and decide what money to spend.
The double 11s have been arranged according to priority -- this time it's bought, it's spent, it's spent.
If you're wondering where all your money is, maybe you'll let the light take a look at you, too.
The trip: plans keep up with changes
If you had a travel plan, you'd probably feel the most: the plan won't keep up.
You spend the night searching, scheduling, booking restaurants, and when the plan is done, friends say there's no time, or if the weather suddenly changes, the whole trip has to be pushed back. The most annoying thing is that I know that I'm reprogramming, but I'm always dragging it down — and then I'm gonna start over.
I've recently tried to turn this over to Shining Light.
The ability of the light to follow and adjust is a surprise to me, given the weather changes or the changes in the schedule, the active window alerts, and the reprogramming of the planned accommodations, routes and rhythms in the light of new developments, so as to minimize the difficulties of change.
My original need for it was this weekend, me and my best friend, and a 4-year-old kid in her family who wanted to go out, drive for less than two hours, not too fast, having to take a nap, we can't even eat spicy, have to be home by 8 o'clock on Sunday night, with a budget of around 400.
In the course of its programme, it asked a number of key questions and then gave a complete schedule: where to go, what to do in the morning, where to eat lunch, where to make it easier for children to take a nap, where to continue in the afternoon, and where to leave at the right time. It also tells me what needs to be brought before we leave on Saturday, and if the weather changes, I will be informed in advance.
And the next day, I told her, "I can't get my best friend to come, I'm alone, I'm rescheduled, the budget is the same, but one can be more flexible and want to add a place."
It was reprogrammed, the original framework was retained, the arrangements involving two persons or children were adjusted, and additional destinations were filled in. One moves without having to set a break, and the rhythm is free to go a little further, and it takes these factors into account.
The context accumulated in the conversation always remembered that you told it what had changed, that it was going to update the relevant parts and not allow you to repeat yourself.
But the plan has not kept pace with the changes, but the changes that have taken place will probably not delay me any longer — because it is no longer entirely my business to reprogramme it.
Muse, run your leg, light. 2.0. Yes.
Socialist Allison Daminger, in a study on the division of labour in the family in 2019, broke down “cognitive work”, which is not easily visible, into four stages: prejudicing needs, looking for options, making decisions, and following up on results. Like cooking and cleaning, they maintain a home.
From this perspective, Personal Agent finds a neglected premise: When we give AI a clear enough order, a considerable part of the work has been done for it.
In APSO, it seems that "supervisor" should be an important dimension of evaluating the individual Agent. In addition to completing the directive, it is also to be seen whether it can share the tasks of identifying problems before and after the order, keeping an eye on them and continuing to organize them.
Almost all of AI is rolling "How to Work for Users" - writing articles, drawings, searching for strategies, generating small applications. But this kind of interaction is still essentially at the tool level of "one input, one output."
Psychic 2.0 Starts to try to drive the AI Assistant to evolve from a passive response to a command, to a single mission, to a proactive identification of needs, to a continuous management.
It took a few days, and I realized I didn't have to look at everything with memory and consciousness.
The value of Personal Agen is reflected in the long, fragmented time before the click "Recognition". With Personal Agent checking information, remembering things, and following up on them in a timely manner, I don't have to worry about anything.
You just have to say what you want once, the rest of the search, comparison, recording and reminder, and it's moving slowly backstage. There's news, it's coming to you; when it's time for you to make up your mind, it puts it in front of you.
Looking back at the "lighting" before, the light is already trying to make ordinary people use their natural language to produce a small, interactive, modified tool to meet each person's specific and small needs.
The logic of making more complex life matters simple, based on individual needs, also continues to the light of this event 2.0. From this perspective, it was also the first batch of Personal Agent products.
The light 2.0 may remind you of the same way you play personal life, but they make me feel very different. If Muse is a "legged" Personal Agent, help me with the mail, the next product. The light 2.0 is the "manager" Personal Agent, which is given more to the little things in life.
It also echoes the lessons of this year's lobster boom: A large number of personnel were deployed without real access and the heat quickly subsided.
Personal Agent patterns are now convergent in the market, and the differences are mainly at the entrance and in the ecology.
The cloud virtual machine, 24 hours of uninterrupted operation, and the second confirmation of sensitive operations have now become an AI Agent ' s industry frame. Spark, for example, asks for permission before paid or e-mailing, and Dots gives high-risk tasks like changing passwords to the user.
Under this standard structure, the core difference between players is the bottom card in their hands: Google relies on mailbox ecology, Meta masters social distribution, Open AI has huge plug ecology...
The ability of Personal Agent to access services, obtain authorization and safely and reliably complete implementation determines whether Agent is sufficiently personal and personal.
Muse is connected to Stripe's Link payment and first gets its shopping security, suggesting that the payment infrastructure will be critical.
It happens that the ecology in which the antslight 2.0 is located is naturally linked to real trading scenarios such as payments, life services and so on, and that these advantages have been translated into the implementation capacity of Agent.
As a result of the tests, the light 2.0 is not only internalized into the payment treasures (billing, credit cards, flowers, loans, travels, memberships, travels, car life, etc.), QQQ-mails, notes, calendars, alarm clocks, text messages, documents, etc. application services and systems services, but also supports the use of tools such as MCP, CLI in an open environment, inter-platforming nails, flying books, drips, Baidu discs, rice houses, ideal/toy/peng/hua car, 12306 and smart hardware equipment.
Personal Agent shouldn't have kept us busy, it had to free us from the busy life. The final decision-making power should, of course, be preserved and the burden of standing by should be set aside.
It's been a long time since I've been able to remember the little things I've been waiting for.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.