就在刚刚,据彭博社报道,字节正在筹备一款类似 Meta Muse 的个人 AI Agent(智能体),希望在新一轮 AI 竞争中追赶对手。
这款机型的目标,是让豆包进一步走出聊天框,成为可以记住使用者需要、主动规划并执行任务的个人助理。

回顾过去的 9 月底,豆包的首页多了一个叫「出行用豆包」的入口。
使用者只需要告诉它想去哪里,就能借助豆包查询机票、购买火车票、叫车,甚至预订酒店。过去需要在几个 App 之间来回切换的操作,如今开始集中到一个聊天框里。

而字节跳动显然还有更大的打算。
从豆包手机,到真正的个人 Agent
字节对个人 Agent 的探索,早于这次彭博社的报道。
据财新此前披露,字节内部早在今年 4 月就启动了代号为「Spell」的项目,主要由豆包手机助手团队负责,探索个人 AI 助理的机型形态。
9 月初,Meta 推出个人智能体 Muse 后,字节进一步加快了项目进度,并计划让 Spell 项目与豆包对话团队合作,将相关能力整合进豆包。
Muse 可以像普通 Chatbot 一样与使用者交流,但它的工作不会随着聊天结束而停止。

使用者可以交代一项需要花费数天甚至数周的任务,让 Muse 自行规划步骤、搜索信息、调用外部服务,遇到需要确认的环节时,再向使用者发来消息。
为了支撑这种运行方式,Meta 还为 Muse 准备了专门的云端虚拟机平台 Muse Secure VM。每个智能体都有独立的执行环境,可以在使用者授权范围内使用浏览器、访问外部服务,持续处理尚未完成的工作。

视角拉回国内,字节实际上已经沿着这条路线摸索了一段时间。
去年 12 月,豆包手机助手工艺预览版亮相。它可以理解手机显示屏内容,借助模拟点击等方式,代替使用者完成部分 App 操作。
今年 9 月,搭载豆包手机助手消费者版的努比亚 NaviX Ultra 正式开售,将这些能力进一步带向普通消费者。

今年 8 月,字节又推出了面向生产力情形的「豆包工作」。与飞书打通之后,它可以读取授权范围内的工作信息,自主拆解任务、调用工具,还能借助虚拟电脑或浏览器完成操作。
到了 10 月,豆包工作还增加了录制个人技能的能力。使用者只需要演示一遍电脑上的操作流程,就可以让 AI 学会重复执行类似任务,甚至在 Mac 锁屏后继续工作。

从手机上的模拟点击,到电脑上的自动化操作,再到接入出行、住宿等生活服务,字节已经为豆包准备了不少可以真正替使用者办事的工具。
Spell 则可能把这些分散在不同机型中的探索,带向一个更完整的个人助理形态。

4.19 亿豆包使用者,和抖音的另一张底牌
工艺之外,字节还有一个相当显著的长处:使用者规模。
QuestMobile 今年 10 月发布的《2026 年三季度 AI 应用市面发展报告》表明,截至 9 月,豆包 App 的月活跃使用者达到 4.19 亿,同比增长 143.6%。
同期,千问和 DeepSeek 的月活分别为 1.58 亿和 1.23 亿。豆包的月活已经超过两者之和,在国内 AI 原生 App 市面占据领先位置。
更值得关注的是使用频率。QuestMobile 数据表明,今年 9 月,豆包月人均使用次数达到 72 次,平均下来每天超过两次。
当其他 AI 机型还在努力争取使用者打开 App 时,豆包已经成为数亿人日常使用的工具。对于字节来说,新机型可以直接建立在这批现有使用者之上,省去了重新培养使用习惯的漫长过程。

但与豆包本身的规模相比,它背后的抖音可能更加值得关注。
QuestMobile 11 月 3 日发布的《中国移动互联网秋季大报告》表明,2026 年 9 月,抖音 App 的月活跃使用者达到 10.29 亿,整个抖音集团的去重使用者规模则达到 12.59 亿。

过去十多年,抖音围绕海量内容和使用者行为,建立起了一套成熟的个性化建议平台。
依据抖音公开的算法备案说明,浏览、搜索、关注、收藏,以及电商和本地生活情形中的互动记录,都是建议平台判断使用者兴趣的关键依据。
一个人最近对什么感兴趣,准备购买什么商品,又在关注哪些消费服务,这些行为都会影响他接下来看到的内容。
抖音已经很擅长预测使用者想看什么。而如果个人 Agent 能在使用者授权下接入相关信息与服务,字节或许就能把这种长处延伸到使用者的下一步行动。
例如,使用者最近在抖音上收藏了不少露营攻略,正在豆包里查询周末的天气和交通。个人 Agent 便有机会把原本零散的操作串联起来,协助筛选营地、制定路线,再借助合作服务完成预订。

过去,抖音的建议平台负责把内容推到使用者眼前,至于看完之后要做什么,还得由使用者自己决定、自己操作。
个人 Agent 则让字节有机会参与后续的流程:从使用者产生兴趣,到提出需要,再到完成一笔交易或安排一项服务。
从「猜你喜欢」到「帮你完成」,字节想争夺的,或许不再只是使用者停留在显示屏前的时间。
本文来源:https://www.ifanr.com/1689016
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Just moments ago, Bloomberg reported that ByteDance is preparing a personal AI Agent similar to Meta Muse, hoping to catch up with competitors in the new round of AI competition.
The goal of this model is to let Doubao move beyond the chat box, becoming a personal assistant that can remember user needs, proactively plan and execute tasks.
Photo | Bloomberg
Looking back to late September, the Doubao homepage added an entry called "Use Doubao for Travel."
Users only need to tell it where they want to go, and it can use Doubao to check flights, buy train tickets, call cars, and even book hotels. Operations that previously required switching between several Apps are now beginning to be concentrated into a single chat box.
Photo | Daily Economic News
And ByteDance clearly has even bigger plans.
From Doubao Phone to a Real Personal Agent
ByteDance's exploration of personal Agents predates this Bloomberg report.
According to previous disclosures by Caixin, ByteDance internally launched a project codenamed "Spell" as early as April this year, mainly led by the Doubao Phone Assistant team, exploring form factors for personal AI assistants.
In early September, after Meta launched the personal agent Muse, ByteDance further accelerated project progress and plans to have the Spell project collaborate with the Doubao dialogue team to integrate related capabilities into Doubao.
Muse can communicate with users like an ordinary Chatbot, but its work does not stop when the chat ends.
Photo | Meta
Users can assign a task that takes days or even weeks, letting Muse plan steps, search information, call external services, and send messages to the user when confirmation is needed.
To support this mode of operation, Meta has also prepared a dedicated cloud virtual machine platform, Muse Secure VM, for Muse. Each agent has an independent execution environment and can use browsers and access external services within the scope authorized by the user, continuously handling unfinished work.
Photo | Meta
Bringing the perspective back to China, ByteDance has actually been exploring this route for some time.
Last December, a technical preview of the Doubao Mobile Assistant debuted. It can understand the content on a phone's display and perform certain App operations on behalf of the user through simulated clicks and other methods.
This September, the Nubia NaviX Ultra, equipped with the consumer version of the Doubao Mobile Assistant, officially went on sale, bringing these capabilities further into the hands of ordinary consumers.
Image | Doubao Mobile Assistant Official Website
This August, ByteDance launched "Doubao Work" for productivity scenarios. After integrating with Feishu, it can read work information within authorized scopes, autonomously break down tasks, invoke tools, and perform operations using a virtual computer or browser.
By October, Doubao Work had added the capability to record personal skills. Users need only demonstrate a computer operation process once to let the AI learn to repeatedly execute similar tasks, even continuing to work after the Mac enters lock screen.
From simulated clicks on mobile phones to automated operations on computers, and further to accessing lifestyle services like travel and accommodation, ByteDance has prepared numerous tools for Doubao that can genuinely act on behalf of users.
Spell, however, may bring these explorations scattered across different devices into a more complete personal assistant form.
419 Million Doubao Users, and Another Trump Card of Douyin
Beyond technology, ByteDance has another significant strength: user scale.
The "2026 Q3 AI Application Market Development Report" released by QuestMobile this October indicates that as of September, Doubao App's monthly active users reached 419 million, a year-on-year increase of 143.6%.
During the same period, the monthly active users of Qianwen and DeepSeek were 158 million and 123 million respectively. Doubao's monthly active users have surpassed the sum of the two, occupying a leading position in the domestic AI-native App market.
Even more noteworthy is the usage frequency. QuestMobile data shows that in September this year, the average monthly usage count per Doubao user reached 72, averaging more than two times per day.
While other AI products are still striving to get users to open their Apps, Doubao has already become a daily tool for hundreds of millions of people. For ByteDance, new products can be built directly upon this existing user base, saving the lengthy process of re-cultivating usage habits.
Image | QuestMobile
But compared to the scale of Doubao itself, the Douyin behind it may be even more worthy of attention.
The "China Mobile Internet Autumn Big Report" released by QuestMobile on November 3rd shows that in September 2026, the monthly active users of the Douyin App reached 1.029 billion, while the deduplicated user scale of the entire Douyin Group reached 1.259 billion.
Photo | QuestMobile
Over the past decade, Douyin has built a mature personalized recommendation platform centered on massive content and user behavior.
According to the algorithm filing instructions published by Douyin, browsing, searching, following, collecting, as well as interaction records in e-commerce and local life scenarios, are key bases for the recommendation platform to determine user interests.
What a person is interested in recently, what products they are preparing to buy, and which consumer services they are paying attention to—all these behaviors will affect the content they see next.
Douyin is already very good at predicting what users want to watch. And if a personal Agent can access relevant information and services with user authorization, ByteDance might be able to extend this strength to the user's next step in action.
For example, a user has recently collected many camping guides on Douyin and is checking the weekend weather and traffic in Doubao. The personal Agent would have the opportunity to string together these originally scattered operations, assist in screening campsites and planning routes, and then complete bookings through partner services.
In the past, Douyin's recommendation platform was responsible for pushing content to users; as for what to do after watching, users had to decide and operate by themselves.
Personal Agents give ByteDance the opportunity to participate in the subsequent process: from users generating interest, to raising needs, and then to completing a transaction or arranging a service.
From "Guessing what you like" to "Helping you complete it," what ByteDance wants to fight for might no longer just be the time users spend in front of the screen.