您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：机票订完后，AI 24小时盯价，发现降价及时提醒；翻看日常账单，AI识别可能闲置的续费订阅，帮你省钱；家长顾不上查看班级群，AI也能梳理老师通知、提取待办事项…
机票订完后，AI 24小时盯价，发现降价及时提醒；翻看日常账单，AI识别可能闲置的续费订阅，帮你省钱；家长顾不上查看班级群，AI也能梳理老师通知、提取待办事项……这些过去需要用户自己发现、记住和安排的生活琐事，如今正开始由AI主动接手。
10月27日，蚂蚁集团发布灵光2.0，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、持续管理事务的演进。
与传统对话式AI及任务型Agent相比，灵光2.0的一大变化，是将AI介入任务的起点进一步前移。 过去，用户需要先发现问题、提出需求，再由AI执行；如今，在获得用户授权后，灵光能够结合已接入的信息与服务，主动识别潜在需求，并发起分析、监测和跟进，并在需要用户关注的节点及时提醒。AI不再只是等待指令的工具，而开始成为能够持续管理事务的助手。
这一变化背后，是灵光2.0对跨会话记忆、长程任务调度和工具连接能力的全面升级。依托独立云主机，灵光能够突破单次对话的限制，在用户退出对话后继续运行任务，并根据新的信息和状态变化动态跟进，让AI从一次性的任务执行转向长期、持续的事务管理。
灵光管理事务的能力正在进入具体的生活场景。
在消费领域，用户授权接入支付宝后，灵光可以梳理消费账单记录，发现可能被遗忘的自动续费订阅，分析消费习惯并提供省钱建议等；在出行领域，机票、酒店预订完成并不意味着AI服务结束，灵光能够持续关注后续价格变化，发现降价机会并及时通知用户；在家庭场景，灵光在授权后可以关注钉钉班级群中的老师通知，整理重要信息、提取待办事项，并完成文档排版等后续工作。
此外，从宠物照护、儿童疫苗接种，到家人纪念日、车辆年检等需要长期记忆和持续跟进的事务，灵光也能够在授权范围内主动识别关键节点、安排后续跟进，减少用户反复查询、提醒和跟进负担。
从单次完成任务，到主动理解并持续推进，意味着AI Agent正在从“任务执行者”向“事务管理者”转变。 要实现这一能力，仅有模型的理解和推理能力还不够，还需要跨应用的信息获取、工具调用以及长时间稳定运行的基础设施支撑。
据介绍，灵光2.0不仅打通了支付宝生态内的数据与服务，还支持在开放环境中调用MCP、CLI等工具，拓展跨平台协同能力。目前，灵光已内置支付宝账单、会员、出行，以及钉钉、飞书、邮箱、米家等服务能力，管理范围正从个人信息与线上事务，延伸至跨平台工作流和智能家居设备控制。
随着智能体走向“事务管理”，其权限边界、数据安全和执行可信度也成为新的挑战。灵光2.0接入了蚂蚁APASS，作为AI时代跨智能体、跨设备、跨应用与服务场景的统一账号和安全管理基础设施，APASS将身份认证、意图安全和隐私保护纳入Agent执行流程，通过授权密钥与Agent分离托管等机制，让智能体在不直接接触授权密钥的情况下执行任务。用户可随时查看、调整或撤销授权，让智能体在授权范围内办事。
据了解，灵光App是蚂蚁集团推出的全模态通用AI助手，于2025年11月18日正式上线，上线6天下载量突破200万。此前，灵光凭借“一句话生成闪应用”、“生成你的世界”等功能，将自然语言交互拓展至应用生成和多模态内容创作，引领了AI助手的全新交互模式。
个人Agent的真正突破，是AI开始拥有时间维度。Chatbot（聊天AI）主要解决当下的问题，而个人Agent需要处理跨越数小时、数天乃至更长周期的事情。它必须记住上下文、感知变化、推进任务，并在必要时重新请求人的判断。
灵光的此次升级也折射出AI助手的下一阶段竞争方向：比起能回答多少问题、完成多少指令，AI能否理解用户的长期需求，在恰当的时机主动介入，并持续把事情管理好，正在成为衡量智能体走进真实生活场景的新维度。
为了您的账户安全，请验证邮箱
请验证您的邮箱
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。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: After the airline ticket is booked, AI is on a 24-hour watch and finds that the price reduction is a timely reminder; looking at the daily bill, AI identifies potential idle renewal subscriptions and saves you money; parents can't look at the class, and AI can comb the teacher's notice and extract to-do...
Once the ticket has been booked, AI is on a 24-hour watch and finds that the price reduction is a timely reminder; looking at the daily bill, AI identifies potential idle renewal subscriptions and saves you money; and the parents are unable to look at the class, and AI is able to comb through the teacher's notice and extract to-do matters... This has in the past required the user to discover, remember and organize himself, and is now beginning to be taken over by AI.
On 27 October, the ants group publish light 2.0 facilitated the evolution of AI assistants from passive response to commands, to individual assignments, to proactive identification of needs and ongoing management.
A significant change from traditional dialogue-based AI and mission-type Agent 2.0 is to move the starting point for AI interventions further forward. In the past, users needed to identify problems, demand and then be executed by AI; today, with the authorization of the user, the light can proactively identify potential needs in conjunction with the information and services that have been accessed, and initiate analysis, monitoring and follow-up, with timely reminders at nodes requiring user attention. AI is no longer just a waiter for instructions, but is beginning to be an assistant who can manage the business on a continuous basis.
Behind this change is the overall upgrade of the light 2.0 capability for cross-session memory, long-range tasking and tool connectivity. Depending on the independent cloud mainframe, the light can break the limits of a single dialogue, continue to run the mission after the user has withdrawn from the dialogue, and follow up with new information and changing state dynamics, moving AI from one-time task execution to long-term, ongoing service management.
The capacity of the light management services is entering a specific life scene.
In the area of consumption, after the user has authorized access to the treasure to pay, the light can combe the record of the consumer bill, discover automatic subscriptions that may be forgotten, analyse consumption habits and provide cost-saving advice, etc.; in the area of travel, the completion of airline tickets, hotel reservations does not mean the end of AI services, the light can follow up on price changes on a continuous basis, discover opportunities for price reductions and inform the user in a timely manner; in the home scene, the light can follow up on the announcement of teachers in the cruising class after the authorization, sort out important information, extract to be done, and complete the document layout etc.
In addition, long-term memory and follow-up services, ranging from pet care, vaccination of children to family days, annual vehicle examination, etc., can also proactively identify key nodes and arrange follow-up within the delegated authority, reducing the user ' s burden of repeated queries, reminders and follow-up.
From a single assignment to a proactive understanding and continued advancement, AI Agent is moving from a “manager” to a “service manager”. To achieve this capacity, model understanding and reasoning alone will not be sufficient, but will also require cross-application access to information, tool call and long-term stable infrastructure.
It was described that light 2.0 not only provided access to data and services within the paying environment, but also supported the use of tools such as MCP and CLI in an open environment to expand cross-platform synergies. At present, the light has built up the capacity to pay for valuable bills, members, travels, and services such as nails, flying books, mailboxes, and rice, which are being managed from personal information and online services to cross-platform workflow and smart home equipment control.
As AI agent moves towards “business management”, its permission border, data security and the credibility of implementation have also become new challenges. Psychic light 2.0 connects ant APASS, which serves as the uniform account number and security management infrastructure for the AI-era trans-AI agent, cross-equipment, cross-application and service scenes, by including identification, intent security and privacy protection in the Agent implementation process, and enabling AI agent to carry out its tasks without direct contact with the Agent ' s authorized key through mechanisms such as the separation of authorized keys from the Agent. Users can see, adjust or revoke the authorization at any time to allow AI agent to do so within its mandate.
Psychic App is known to be an all-modular General AI assistant launched by the ants group, officially on line on 18 November 2025, with a six-day download exceeding 2 million. Prior to that, the light led to a whole new interactive mode of AI's assistant by expanding natural language interaction to application generation and multi-modular content creation, using the functions of "one sentence to generate flash applications" and "the world to produce you".
The real breakthrough of the individual Agent was that AI started having time dimensions. Chatbot (Chattering AI) deals primarily with current problems, while the individual Agent has to deal with events that span hours, days and even longer. It must remember context, perceive change, advance the task and, if necessary, renew the judgement of the person requesting it.
This upgrade also reflects the next phase of the competition of AI’s assistant: whether AI understands the long-term needs of users, intervenes at the right time and continues to manage things well, compared to how many questions can be answered and how many instructions can be completed, is becoming a new dimension of measuring AI agent’s entry into real life.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.