智东西（公众号：zhidxcom）
作者 | 毕伟豪
编辑｜漠影
近几月，Muse的出现，带火了Personal Agent，也让整个行业开始持续思考Personal Agent这一产品形态。
但Muse的成功，显然不是引入形式就能复制的，依托长期记忆和云端虚拟机，让Agent能够进行持续的事务管理，这套打法思路非常清楚，但想要真正实现，显然还有很多“形式”之外的“内容”。
10月27日，蚂蚁集团发布了灵光2.0版本，将产品定位从强调对话和应用生成的AI助手，进一步推到“能管事的个人助理”。
灵光2.0版本从形式上看，是“中国版”Muse，但从内容上，更像是“中国化“的Muse。
因为蚂蚁集团并不是简单提供了持续执行能力，灵光2.0将这项管事能力与支付宝、钉钉、飞书、米家等平台的各项功能进行了适配，试图让Personal Agent与中国用户实际使用的服务网络深度融合。
智东西也对灵光2.0进行了深度体验，切身感受了一下这份持续管理个人事务的能力。
一、实测灵光2.0：租房、比价、收支盘点
对灵光2.0的体验，我没有特意去思考用什么提示词、或者测试哪些任务，就是随着个人日常使用慢慢感受，也因此，所有的提示词都是普通用户的视角。
新版本灵光的页面和之前有了很大变化，重点是页面中间的这些需要灵光管理的事务，以及下面一个“要我管”的功能区域，任务完成和变动后会有消息弹窗提示。
用户只需要点击左下角灵光，就可以通过对话来将任务交给它了。
拿到灵光2.0的第一时间，我最先想到的，是让它帮我找找房子，找房子的任务往往持续数月，任务型Agent无法满足长时间线作战的需求，正是检验灵光2.0的最佳场景。
我输入了一段提示词，希望能在明年2月份房租到期的时候，找到一个合适的新房子，提示词如下：
我想在明年2月份换房子，帮我盯一下北京市北六环外的房源，离清河近一些的，小区或者自建房都可以，价格在1500到2200之间。
最开始的提示词并不完善，灵光根据需求给出了一版方案，但也提示我缺少很多信息。通过多轮对话交互，灵光帮我敲定了一份租房计划。
这份计划的重点，是“盯梢”。灵光2.0会按照既定的时间安排进行搜集和汇报，然后给出房源信息，不断追踪任务进展，直到完成。
随后，我让灵光来帮忙记账，授予支付宝权限后，输入了我的收入和各项开支，让它每个月给我盘点一下能存多少钱，如何优化开支情况。
灵光2.0很快就构建好了盘点任务，并通过我授权的支付宝信息，对我日常的支出和收入进行了非常细化的统计，能看到，保险支出、高速费、房租、取暖费等等都一一列举了。
对于日常使用支付宝较多的用户，这个记账助手还是非常好用的。
双十一购物，也是一个持续性的任务。灵光2.0可以结合购物平台信息，进行商品比价和购置规划。
比如我输入了买衣服的需求，让灵光帮我推荐，提示词如下：
双十一我想买些衣服，秋冬的卫衣、短款羽绒服、工装裤、板鞋这些吧，帮我比比价格，计算下优惠的购物方案，然后有变化随时调整和通知我。
灵光立刻给出了方案，还在下面提示我补充预算、平台等信息。
现在灵光正在帮我盯着平台上的商品价格，它还会关注各类服饰的直播专场，有低价潮出现时会提醒我。
除此之外，灵光2.0还在帮我盯着Switch游戏的价格，比如塞尔达、马里奥系列的卡带以及数字版。这些东西的价格变动非常快，因此需要灵光2.0能在低价出现的时候及时提醒我。
我还让灵光2.0帮我统计每天的消费开支，并在超过50元时提醒哪里花超了。
从这些测试中可以看出，灵光2.0不仅在持续管理个人事务上具备不俗的能力，更是将这份能力与一些生活中常用的服务深度融合。
最重要的是，这些任务是跑在云端的，可持续，不会因为关掉应用就停止。同时，云端独立主机也为任务执行提供了干净的环境。
Personal Agent，也在大厂的生态下，展现出了中国本土化的样子。
二、Muse之外，灵光探路中国式Personal Agent
对话式AI解决单次提问，Agent面向复杂任务处理。两者还有一个共同点：事情交给AI之后，用户仍然要自己跟进来完成任务。
Personal Agent要做的，是接手这些后续工作，让Agent记住目标、保留任务进程，在出现新信息时继续推进，并在关键节点提醒用户。
这些要求落到灵光2.0的产品上，是这几项能力的组合：跨会话记忆、长程任务调度、独立云主机以及相关的工具连接。
记忆负责保留任务背景，调度负责推动任务继续进行，工具负责获取新信息、调用外部服务，独立云主机则保障用户退出对话后的持续运行。
▲灵光2.0盯金价，每小时监控大盘并实时更新
但机制以外，服务生态是完成任务的核心支撑，Muse爆火让行业看到Personal Agent的潜力，但产品形态可借鉴，落地所依赖的服务环境却不能简单照搬。
Muse于今年9月8日上线后，在美国提供iOS、Android和网页端服务，其公开的商业合作伙伴包括Stripe、Shopify、Expedia等，覆盖支付、电商、旅行预订和零售等场景。
对海外用户而言，订单、行程和消费信息往往通过邮箱、网站获取，后续操作也能借助浏览器完成。Agent通过连接邮箱、操作网页，就有机会将分散的信息串联起来，持续推进任务。
但这套思路落到中国市场，会有一些水土不服。中国用户的日常需求散落在各大APP上，订票、消费、支付、办公，各自有成熟的App和服务体系，信息和操作入口未必相通。
即使Agent能够理解需求、规划步骤，如果无法顺畅获取这些信息、调用相应服务，任务仍可能卡在应用之间，需要用户自己接手。
因此，Personal Agent进入中国市场，关键不只是复现持续记忆和自主执行的产品形态，还要找到连接本土服务生态的方式。这也是灵光2.0将支付宝、钉钉、飞书等服务纳入工具体系，尝试让AI从单次执行走向持续管事的原因。
具体而言，灵光2.0接入支付宝账单、会员、出行等服务，车生活一侧覆盖停车、加油和充电，同时支持QQ邮箱、滴滴、钉钉、飞书、百度网盘、米家等平台，开发者还可以通过MCP、CLI等方式扩展灵光的能力。
▲灵光2.0接入各大应用、查找汽车充电站
此外，产品还提供独立电脑和浏览器，同样可以通过网页操作处理任务；其开放能力也可延伸至理想、蔚来、小鹏、华为汽车、运动手环、体脂秤和12306等设备与服务。
这些连接不只是增加了可调用的工具数量，更决定了Agent能够接触哪些信息、推进哪些环节，也决定了“持续管事”能在多大程度上进入中国用户的日常生活。
三、服务生态与信任机制，让Personal Agent落地生根
从灵光1.0到2.0，产品重心也从内容创作延伸到了事务管理。
2025年11月上线的灵光1.0，主推全模态交互、灵光闪应用和灵光开眼，上线6天下载量突破200万。其中，灵光闪应用可通过一句话在约30秒内生成可交互、可编辑、可分享的小应用，灵光开眼则侧重多模态识别与理解。彼时，灵光主要探索如何降低AI使用和应用生成的门槛。
2.0的灵光开始尝试接手那些无法一次完成、需要持续跟进的事务，试图让AI记住目标、持续推进任务，并根据变化及时跟进。
而当AI能长期接触个人信息、调用外部工具、处理涉及资金和重要事务的操作，权限边界与执行可信度就不可回避。蚂蚁在支付安全和风险控制方面的积累，为可信智能体的探索提供了基础。
支付宝2004年独立运营，至今22年，支付安全、身份认证、授权管理这套能力，是在处理资金类业务的过程里一点点磨炼出来的。
在授权方面，灵光2.0在接入其他应用时要用户明确授权，用户可以随时查看、调整或撤销授权，智能体的动作被限制在授权范围内。
▲连接钉钉需要明确授权，通过APASS连接
信任之外，蚂蚁还有一张深入日常生活的服务网络。二十余年间，支付宝连接起超10亿用户、8000万商家，上万项服务覆盖消费、出行、缴费与政务民生。这些积累，为AI承接个人事务提供了具体的去处。
灵光2.0的探索，让这张网络有了新的想象空间。持续记忆与任务跟进，若能与成熟的商家和生活服务衔接，用户的一次托付，就有机会从计划走向预约、下单与办理。
作为蚂蚁集团探索Personal Agent的前沿应用，灵光若能验证持续记忆、任务调度与可信执行等能力，并将其复用于更多高频场景，未来就有望依托蚂蚁的服务生态与支付基础设施，触达更多用户。
结语：Personal Agent竞争升温，蚂蚁持续探索本土路线
Muse让行业加深了对Personal Agent的想象，灵光2.0则尝试回答另一个问题：这种产品形态怎么与中国用户的实际生活连接。
而要“中国化”，就需要深度结合本土应用、工具、服务生态，同时建立可信基础设施。
对蚂蚁而言，灵光2.0是一次前沿产品探索，也为支付宝生态中的服务能力如何与个人智能体结合提供了新的实践方向。
Personal Agent尚无定式。率先出发，可以争取先机；能否走得更远，终究要看产品能否经得起真实生活的检验。灵光2.0，正是蚂蚁集团沿着这条路线展开的一次深度求解。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
The author.
Edit Indigo
In recent months, the emergence of Muse has led to the burning of Personal Agent, which has also allowed the entire industry to begin to think continuously about its product pattern.
But it is clear that the success of Muse is not replicable by introducing a form, and it is clear that there are many more “forms” than “forms” to actually achieve it, thanks to long-term memory and cloud virtual machines, which allow Agent to manage his affairs on a continuous basis.
On 27 October, the ant group publish produced version 2.0 of the light, which further moved the product from the AI assistant, who emphasized dialogue and application generation, to the “personal assistant who can manage”.
In formal terms, it's the Chinese version of Muse, but more like Muse in content.
Because the ants group does not simply provide a continuous capacity for implementation, the light 2.0 adapts this management capacity to the functions of platforms such as paying for treasures, nails, flying books, rice houses, etc., in an attempt to integrate Personal Agent in depth with the service networks actually used by Chinese users.
Intellectuals also have a profound experience of the light 2.0 and feel the ability to manage their personal affairs on an ongoing basis.
I. Factual light 2.0: Rental, value, income and expenditure balance
I didn't mean to think about what prompt was going to be, or what was going to be tested, because I was slowly feeling with the daily use of a person, and therefore all prompt was an ordinary user's perspective.
The new version of the light page has changed considerably before, focusing on these matters in the middle of the page that require light management, as well as the following functional area of "I want to manage" with a message window alert when the task is completed and changed.
All users need to do is click on the light in the lower left corner so that they can give it the task through dialogue.
The first thing I thought about when I got a light of 2.0 was that it helped me find a house, which often lasted for months, and the mission-type Agent was unable to meet the need for a long line of action, which was the best scenario for testing the light of 2.0.
I entered a section of prompt, hoping to find a suitable new house by the time the rent expires next February. prompt is as follows:
I'd like to change houses in February next year, and I'd like to keep an eye on the source of houses in Beijing's six-ring area, near the river, either in the district or on my own, at between 1,500 and 2,200.
The first prompt is not perfect, and the light gives a version of it based on demand, but it also points to a lack of information. Through interactive rounds of dialogue, the light helped me finalize a rental plan.
The focus of the project is on “stalking”. The light 2.0 is collected and reported according to the established schedule, then gives information on the source of the house and keeps track of the progress of the mission until it is completed.
I then asked the light to help with the bookkeeping and the payment of the treasure permission, and I entered my income and expenses so that I could take a monthly inventory of the amount of money that could be saved and how to optimize the expenditure.
The light 2.0 mission was quickly completed and, through my authorized payment information, a very detailed accounting of my daily expenditure and income can be seen in the list of insurance expenses, high-speed charges, rents, heating, etc.
This bookkeeping assistant is also very useful for users who pay more money on a daily basis.
Two-one shopping is also an ongoing task. Psychic 2.0 allows the planning of commodity prices and acquisitions in conjunction with the information on the shopping platform.
For example, I've entered the demand for clothes to be recommended by Shining Light.
I'd like to buy some clothes, some clothes for the autumn and winter, some shorts, some pants, some shoes, to help me compare the price, to calculate the preferential shopping options, and to adjust and inform me of any changes.
The light immediately gave me a plan and further updates on the budget, platform, etc.
Now the light is helping me keep an eye on the prices of commodities on the platform, and it's also watching live shows of all kinds of clothing, and it's reminding me when low prices come up.
Besides that, Psychic 2.0 is helping me to keep an eye on the prices of Switch games, such as Zelda, Mario series tapes and digital versions. The prices of these things are moving very fast, so the light 2.0 is needed to warn me when the price is low.
I've also asked Spiritlight 2.0 to help me measure daily consumption spending and remind me of spending more than $50.
From these tests, it can be seen that the light 2.0 not only has a remarkable ability to manage personal affairs on a continuous basis, but also to integrate it with the depth of services that are commonly used in some lives.
Above all, these tasks are run over the clouds, sustainable and do not stop because they are shut down. At the same time, the cloudy independent mainframe provides a clean environment for mandate implementation.
Personal Agent, also in the ecology of large plants, has shown China's indigenousization.
II. Chinese-style Personal Agent outside Muse
Dialogue AI addresses single questions and Agent addresses complex tasks. There's also one thing in common: once things have been given to AI, the users still have to follow in and do their job.
What should be done is to take over the follow-up, to remind Agent of the goal, to keep the task process alive, to move forward as new information emerges and to alert users at key nodes.
These requirements, which are placed on the luminous 2.0 product, are a combination of these capabilities: cross-conference memory, long-range tasking, stand-alone cloud hosts and associated tool connections.
Memory is responsible for keeping the background to the mission, movement control is responsible for facilitating the continuation of the mission, tools are responsible for obtaining new information and calling external services, and independent cloud hosts guarantee the continuity of the user after withdrawal from the dialogue.
光 2.0 on gold price, monitor disks every hour and update them in real time
Beyond the mechanism, however, the service ecology is at the heart of the task, and the Muse fire has allowed the industry to see the potential of Personal Agent, but product patterns can be replicated and the service environment on which it depends cannot be easily replicated.
After coming online on 8 September this year, Muse provided iOS, Android and web-end services in the United States, with open business partners including Stripe, Shopify, Expedia and others, covering such scenarios as payments, electricians, travel reservations and retail sales.
For overseas users, orders, itineraries and consumer information are often obtained through mailboxes, websites and follow-up can be done with browsers. By connecting mailboxes and operating web pages, Agent has the opportunity to link decentralized information and continue to advance the mission.
But this is going to fall into the Chinese market, and there's going to be a bit of land and water. The day-to-day needs of Chinese users are scattered over the major APPs, which book, consume, pay, work, have well-established App and service systems, and access to information and operations is not necessarily compatible.
Even if Agent is able to understand the needs and plan the steps, if access to this information and the corresponding services is not readily available, the task may still be stuck between applications, requiring the user to take over.
Thus, the key to the entry of Personal Agent into the Chinese market is not only the re-emergence of continuous memory and self-implemented product patterns, but also the identification of ways to connect home-grown service ecology. This is also the reason why Shining 2.0 integrates services such as the payment of treasures, nails, flying books into the tool system in an attempt to move AI from single execution to continuous management.
Specifically, the light 2.0 is connected to services such as the payment of precious bills, membership, travel, the side of the car covers parking, refuelling and charging, while supporting platforms such as QQQ mailboxes, drops, nails, flybooks, Baidu web discs, rice houses, etc. The developer can also expand the power of the light through MCP, CLI etc.
Violin 2.0 access to major applications and search for vehicle charging stations
In addition, products provide stand-alone computers and browsers, which can also be processed through web-based operations; their openness can extend to such devices and services as ideal, incipient, peng, Chinese automobiles, sports bracelets, fat scales and 12306.
These connections not only increase the number of tools available, but also determine what Agent can access and move forward, as well as the extent to which “continuing managers” can enter the daily lives of Chinese users.
III. Service Ecology and Trust Mechanisms for Sustainable Development Root
The focus of the product also extends from the creation of content to the management of the service, from luminaries 1.0 to 2.0.
In November 2025, the light was 1.0 and the main thrust was full-modular interaction, light-light applications and light-opening, with a six-day download exceeding 2 million. Of these, light flash applications can generate interactive, editable and shared small applications in about 30 seconds in one sentence, with light-opening focused on multimodular identification and understanding. In other cases, the light primarily explores ways to lower the threshold for AI use and application generation.
The 2.0 light begins to attempt to take over matters that cannot be completed at once and that require sustained follow-up, trying to keep AI in mind its objectives, moving forward on a continuous basis and following up on changes in a timely manner.
The credibility of permission borders and implementation cannot be avoided when AI has permanent access to personal information, calls external tools and handles operations involving funds and important matters. The accumulation of ants in payment safety and risk control provides a basis for the exploration of credible AI agent.
The payment chain operated independently in 2004 and has been in place for 22 years now, with a small degree of refinement in the processing of fund-type operations.
In terms of delegation of authority, the light 2.0 requires explicit user authorization for access to other applications, which can be viewed, adjusted or withdrawn at any time, and the movement of AI agent is restricted to the authorization.
Accompanying nails need clear authorization to connect through APASS
In addition to trust, an ant has a network of services that go into daily life. Over the past two decades, payment treasures have been connected to more than 1 billion users, 80 million businesses, and tens of thousands of services have covered consumption, travel, payment and public life. These accumulations have provided AI with specific access to personal affairs.
The exploration of the light 2.0 gives the network a new imagination. Continuous memory and mission follow-up, and if linked to mature business and life services, a user ' s one-time deposit offers an opportunity to move from a plan to an appointment, order and process.
As an ant group to explore the front-line applications of Personal Agent, if the light can validate the ability to sustain memory, mission movement and credible execution and re-use it for more high-frequency scenarios, it is expected that the service ecology and payment infrastructure of the ant will reach more users in the future.
Concluding remarks: Personal Agent competes for warmth and ants continue to explore indigenous routes
Muse deepened the industry's imagination of Personal Agent, while Shining 2.0 tried to answer another question: How does this product form relate to the actual lives of Chinese users?
“Chinaization”, on the other hand, requires a deep combination of indigenous applications, tools and service ecology, along with a credible infrastructure.
For ants, luminous light 2.0 is a front-line product exploration and provides a new practical direction on how service capacity in the precious ecology can be combined with individual AI agent.
Personal Agent has no fixed pattern. The first step depends on whether the product can withstand real life tests. Spiritlight 2.0, which is the depth of the ants' group along this route.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.