文 | Tech商业
10月29日，字节跳动旗下的AI应用开发平台“扣子”（Coze）正式发布《扣子企业版下线公告》。公告宣布，企业标准版与企业旗舰版将于2026年11月30日起停止接受新客户购买，存量套餐最晚于2027年12月31日全面下线，超期未使用部分将自动退款。
消息一出，业内哗然。在2025年的市场统计中，扣子还曾以19.3%的市场份额稳居中国智能体开发平台公有云市场第一。但在庞大的字节帝国里，市占率第一的光环，并不能保证一个产品线的独立生存权。
扣子企业版的退场，并非一场突如其来的意外，而是一场自上而下的战略裁决。它标志着字节跳动的AI战略正在全面走向“收拢”，也向外界展示了字节内部资源调配的底层逻辑：在“粗主干”战略面前，任何未能形成爆发式商业闭环、且与主干业务重叠的边缘产品，都随时可能面临被降级、整合甚至关停的命运。
组织架构的先兆
扣子企业版的命运，在2026年夏天就已经写好。
将时间线拉回2026年7月底，字节跳动进行了一轮极其低调但影响深远的组织架构调整。飞书产品团队与豆包产品团队正式整合，由豆包负责人统一负责；同时，飞书GTM（Go-to-Market）团队划归火山引擎。紧接着的8月24日，TRAE和扣子团队被整体并入豆包体系，办公相关能力与豆包深度融合，TRAE IDE和CLI转为豆包品牌下的编程产品线。紧接着，字节于8月25日正式推出独立的AI办公产品“豆包工作”。
这一系列密集的动作，是字节AI战略收拢的前奏。在随后的内部全员会上，字节跳动CEO梁汝波明确提出了“高度优先，粗主干，优化长期”的业务原则，并希望豆包能成为如同抖音一般的主干业务。
在“粗主干”的逻辑下，字节的资源正在向极少数的核心目标倾斜。过去，字节擅长“赛马机制”，允许各个团队在细分赛道自由探索，跑出来的获得全量资源。但在AI时代，底层大模型的训练成本、算力消耗以及企业级服务的交付成本极其高昂。当AI竞赛进入深水区，大厂必须集中优势兵力打歼灭战，再也无法承受多条战线同时进行高昂的试错。
扣子自2024年上线以来，凭借零代码搭建Bot和一键发布到飞书、微信、抖音等生态的优势，积累了海量开发者。2025年下半年，扣子试图切入企业办公提效，推出了企业标准版和旗舰版。
但这恰恰触碰了字节当下的战略红线。字节高层的判断非常清晰：办公Agent的仗，必须由“豆包工作”这个主干品牌和火山引擎的私有化底座去打。扣子作为一个独立的办公企业SaaS套餐，不仅与“豆包工作”的功能边界日益模糊，甚至形成了内部同质化竞争。在字节的版图中，办公场景的未来属于豆包，底层算力与企业级交付的未来属于火山引擎。而夹在中间的扣子企业版，失去了独立存在的组织土壤。
第一名的商业困境
让扣子企业版失去独立存在资格的，不仅是组织的意志，还有冰冷的商业数据。
虽然扣子在2025年拿下了中国智能体开发平台公有云市场19.3%的份额，位列第一。但同一份测算显示，该年智能体开发平台公有云市场全年规模仅约2.5亿元。按此测算，扣子的平台产品收入约0.48亿元，不足五千万元（该数字为基于市场规模的估算，不含模型调用与云资源，并非字节官方财务数据）。
这是一个极其刺眼的数据。在互联网大厂的评价体系里，一个全年营收不足五千万、且需要持续投入大量研发和算力资源的平台产品，一旦失去独立增长的爆发力，就会迅速从“战略级资产”降级为“边缘耗材”。
2025年下半年，企业级AI Agent市场迎来了激烈的抢滩战。腾讯WorkBuddy凭借企业微信这一国内渗透率最高的B端组织连接器，打通了身份、通讯录、聊天和文档，迅速覆盖了半数中国500强企业；阿里则依托钉钉生态持续施压，并在此后整合千问办公能力，在B端阵地步步紧逼。
在这样的大厂ToB肉搏战中，扣子企业版并没有发挥出足够的火力。它既缺乏像企业微信那样具有绝对统治力的B端连接器作为护城河，在私有化部署、数据合规与企业级治理的硬核能力上，又不如火山引擎的底座扎实。对于字节而言，用一个独立的企业SaaS套餐去和腾讯、阿里打消耗战，投入产出比极不划算。
于是，字节选择了最符合自身利益的解法：不再让扣子以独立的SaaS套餐形式去试错，而是将其能力拆解。办公属性融入“豆包工作”，开发属性回归纯粹的开发者与垂直场景平台。集中优势兵力，用主干产品去打大厂间的正面战争。
分兵作战与主力会战
扣子企业版的下架，是字节AI整体战略从“分兵作战”转向“主力会战”的缩影。
梳理字节在AI领域的布局，可以清晰地看到三条主线：面向C端的豆包App，面向B端办公的飞书及即将推出的“豆包工作”，以及面向底层基础设施与ToB交付的火山引擎。而扣子，原本被寄予厚望成为连接开发者、企业与字节生态的“万能胶水”。
但“万能”往往意味着“不精”。在AI发展的早期，行业需要低门槛的构建平台来教育和繁荣生态。但当行业进入商业化落地阶段，企业客户需要的不再是一个可以随便拼凑插件的简单平台，而是深度嵌入业务流、具备极高安全性和可观测性的数字员工。
这正是“豆包工作”与火山引擎的组合可以提供的。飞书团队多年的B端产品经验、豆包在模型能力上的持续迭代，加上火山引擎在云原生与私有化部署上的基础设施，构成了字节ToB战略的完整拼图。
在这个新拼图里，扣子企业版的位置变得十分尴尬。如果继续保留其独立的SaaS套餐，不仅会造成内部资源重复投入，还会给客户带来认知混乱——客户到底是该买扣子企业版，还是等“豆包工作”？对于以“大力出奇迹”和“极致效率”著称的字节来说，这种内部的摩擦与内耗是不可接受的。
因此，扣子企业版的退场，本质上是一次内部的资源重组。字节通过组织手段，强行理顺了产品线，确保所有力量汇聚于“豆包”和“火山引擎”这两个最核心的支点上。
大厂产品的“耗材论”
扣子企业版的命运，也是大厂内部赛马机制的典型终局。
在科技圈，有一个越来越明显的趋势：大厂AI产品的存续周期，甚至比创业公司的产品还要短。仅2025年，全球范围内就有大量AI工具停止服务；OpenAI自身也有多个AI工具在不足一年内停止独立服务，例如ChatGPT Atlas浏览器和可视化Agent搭建工具Agent Builder均在上线后不久即宣布下线。
这背后的本质不是大厂做不好AI，而是大厂对待AI产品的逻辑变了：产品不再仅仅是用来赚钱的独立实体，更是验证战略方向的“耗材”。
大厂产品的短周期，往往是“功能被吸收”。无论是扣子企业版的办公能力被整合进“豆包工作”，还是OpenAI的各类小工具被拆进ChatGPT主端。从产品品牌的视角看，它们退场了；但从能力的视角看，它们换了一个更强大、更核心的容器继续存在。对大厂而言，这是资源再配置的必要代价，也是保持组织战斗力的必然选择。
在大厂的机器里，没有哪个具体的产品是不可替代的，只有公司的战略目标不容动摇。当一个产品无法独立支撑起百亿级的想象空间，或者当它的存在影响了主干业务的推进时，被收编、被下线就是它必然的宿命。字节对扣子企业版的果断切割，展现了这家巨头在战略聚焦上的冷酷与高效。
创业公司的生态位
扣子企业版的退场，给整个AI Agent赛道留下了深刻的启示。
对于大厂内部的产品团队而言，判断一个产品能否长久独立存活，不仅要看外部用户的增长，更要看它是否处在公司的“主干”上。一旦与核心战略产生重叠或冲突，再高的市占率也可能在一夜之间被收回。
而对于创业公司和中小厂商来说，大厂的战略收拢恰恰是一个巨大的机会。大厂战略收拢，意味着它们放弃了那些“又苦又累”、需要极度贴合特定行业流程、且无法形成平台级垄断的细分市场。
例如开源AI平台Dify，它于2026年完成3000万美元融资，并选择了一条大厂不愿意走的重活路线：把控制权交给客户，赚取企业为“数据不出门”和“系统可治理”买单的订阅费。Dify不去争夺海量的个人开发者，而是直接切入对数据安全极度敏感的中大型企业，让客户把系统装进自己的机房里。
避开大厂的“主干”航道，去做那些大厂不愿意做的重活、脏活，去服务那些大厂标准化产品无法覆盖的长尾需求，反而能建立起大厂无法轻易拔除的护城河。在大厂追求“粗主干”的时代，垂直与定制，就是创业公司最好的壁垒。
Agent赛道还远未到终局。当AI Agent的计价方式从“软件订阅”转向“任务交付”，千亿级别的市场空间才刚刚打开。
字节的收拢与扣子企业版的退场，只是这场漫长马拉松中的第一个路标。它提醒所有从业者：在巨头的版图里，产品永远服务于战略，随时可能被按下停止键；而在独立闭环里，产品必须服务于客户，只有解决具体的痛点，才能掌握自己的存续权。在AI这场残酷的淘汰赛中，认清自己的生态位，比盲目追求大而全更重要。
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。
Man, Tech Business
On 29 October, the official publish Enterprise Downline Bulletin for Coze, an AI application development platform under the flag of byte, jumped. The bulletin announced that the enterprise standard version and the enterprise flagship version would cease to accept the purchase of new customers as of 30 November 2026, that stock packages would be fully closed down by 31 December 2027 and that the unused portion would be automatically refunded.
As soon as the news came out, the industry was in shock. In the 2025 market statistics, Coze was also at the top of the public cloud market on China ' s AI agent development platform with a market share of 19.3 per cent. In the vast empire of byte, however, the halo of the city ' s highest share does not guarantee the right to an independent existence of a product line.
The exit of the Coze enterprise version was not an unexpected surprise, but a top-down strategic decision. It marks the beginning of the byte jump-starting AI strategy towards “close-up” and shows the outside world the bottom logic of the byte internal resource allocation: In the face of the strategy of “grave backbone”, any marginal product that fails to form an explosive commercial closed ring and overlaps with the backbone may face the risk of being downgraded, integrated or even halted.
A precursor to the organizational structure
The fate of the Coze corporate version was written in the summer of 2026.
By pulling back the timeline to the end of July 2026, the byte beat had a very low-key but far-reaching organizational restructuring. The Flying Books Team was formally integrated with the Doubao Team, under the responsibility of the head of Doubao, while the Flying Book GTM (Go-to-Market) Team was assigned to Volcengine. Immediately thereafter, on 24 August, the TRAE and Coze teams were integrated into the Doubao system, office-related capacity was deeply integrated with Doubao, and TRAE IDE and CLI were converted into programming product lines under the Doubao brand. Immediately thereafter, on 25 August, byte officially launched the independent AI office product “Doubao Job”.
This series of intensive moves is a prelude to the byte AI strategic convergence. At the subsequent in-house staff meeting, byte beat CEO Lubo clearly set out the operational principles of “high priority, heavy backbone, optimization of the long term” and expressed the hope that Doubao would become the backbone of business like Douyin.
Under the logic of “bold backbone”, byte resources are tilting towards very few core objectives. In the past, byte had been good at “racing mechanisms”, allowing teams to explore freely in a subdivision of the track and to come out with full resources. In the AI era, however, the training costs of the bottom LLM were extremely high, the cost of computing and the cost of delivering enterprise-level services. When the AI competition enters the deep water zone, large factories must concentrate on the destruction of their strengths and can no longer withstand the high risks of simultaneous trials on multiple fronts.
Since it went online in 2024, Coze has accumulated large-volume developers with the ecological advantage of establishing Bot and one key publish to fly books, WeChat and Douyin. In the second half of 2025, Coze attempted to gain access to the enterprise office by launching the enterprise standard and flagship versions.
But this is exactly the strategic red line that touches byte now. The high-level judgement of byte is clear: the battle for Agent must be fought by the “Doubao Job” main brand and the privatization base of Volcengine. Coze, an independent office enterprise, SaaS, not only has the functional boundary with “Doubao jobs” become increasingly blurred, but it has even created internal homogenization competition. In the byte layout, the future of the office scene belonged to Doubao and the future of the bottom computing and enterprise-level delivery to Volcengine. And the Coze enterprise version in the middle lost its independent organizational soil.
Number one business dilemma.
It is not only the will of the organization, but also cold business data that has disqualified the Coze corporate version from independent existence.
Although Coze took the 19.3 per cent share of the public cloud market in China ' s AI agent development platform in 2025, it ranked first. However, the same measurement showed that the public cloud market for the AI agent development platform was only about $250 million a year. On this basis, Coze ' s platform products generated about $0.48 million, or less than $$50 million (the figure is estimated on the basis of market size and does not include model deployments and cloud resources and is not official byte financial data).
It's an extremely impressive figure. In the evaluation system of large Internet plants, a platform product that receives less than 50 million dollars a year and requires a sustained investment of significant research and development and computing resources will quickly be downgraded from “strategic-level assets” to “marginal consumption material” once it loses its capacity for independent growth.
In the second half of 2025, the company-level AI Agent market witnessed a heavy beach war. TencentWorkBuddy, with the company WeChat, the B-end organizational link with the highest domestic penetration rate, managed to access identity, correspondence, chat and documentation, quickly covering half of the top 500 Chinese enterprises, while Ali continued to apply pressure by nailing to the ecology and, since then, to consolidate his office capacity and step forward in position B.
Coze did not have sufficient firepower in the Coze Enterprise during the ToB meat fight at such a large plant. It lacks an absolutely dominant B-end connector like WeChat as a moat, a hard-core capability for privatization deployment, data compliance and enterprise-level governance, and less solid than the base of Volcengine. For byte, the use of an independent business, SaaS, to fight with Tencent, Ali, and the input output was extremely uneconomical.
Thus, byte chose the solution best suited to its own interests: instead of allowing Coze to try it in the form of an independent SaaS package, it disintegrates its capacity. Office properties are integrated into “Doubao jobs” and development properties revert to purely developers and vertical scene platforms. Focus on the strength of the force, using the main product to wage a positive war between large plants.
We'll fight with the main forces.
The lower shelf of the Coze enterprise version is a microcosm of the overall strategy of byte AI, which has shifted from “scattered combat” to “main combat”.
The layout of byte in the AI field is clearly visible in three main lines: DoubaoApp for the C side, flybooks for the B side and the forthcoming “Doubao Job” for the B side, and Volcengine for the ground infrastructure and ToB delivery. Coze, on the other hand, had been placed in high hopes of becoming “one-size-fits-all glue” linking developers, businesses and byte ecology.
But “everything” often means “impure”. In the early stages of AI development, industries needed low-threshold platforms to educate and flourish. However, as the industry enters the commercialization phase, what business clients need is no longer a simple platform that can simply collide plugins, but a digital workforce that is deeply embedded in business streams and has a high degree of safety and visibility.
This is the combination of Doubao and Volcengine. The years of B-end product experience of the flying book team, the continuous iterative evolution of Doubao in modelling capabilities, together with the infrastructure of Volcengine in the deployment of clouds and privatization, constitute a complete puzzle of the bytetoB strategy.
In this new puzzle, the position of Coze Enterprise has become very awkward. If its separate SaaS package were to be maintained, it would not only duplicate internal resources, but also create confusion among clients - I'm sorry. Should the client buy the Coze corporate version or wait for the Doubao job? This internal friction and internal consumption is unacceptable for byte, known for its “great miracles” and “extreme efficiency”.
Thus, the exit of the Coze enterprise version was essentially an internal reorganization of resources. By organizing itself, byte has forced the harmonization of product lines to ensure that all forces converge on the two core points of “Doubao” and “Volcengine”.
Large plant products' “depletion material theory”
The fate of the Coze enterprise version is also the typical end of the horse racing mechanism within the plant.
In the science and technology community, there is a growing trend towards the survival of large plant AI products, even shorter than those of start-ups. In 2025 alone, a large number of AI tools ceased service globally; Open AI itself had several AI tools to stop independent service in less than one year, such as the ChatGPT Atlas browser and the visualization of Agent set-up tool, Agent Builder, announced offline shortly after it had been online.
The essence behind this is not that the large plant is not doing AI well, but rather that the logic of the large plant's treatment of the AI product has changed: the product is no longer just an independent entity that is used to make money, but is a “consumption material” to validate strategic direction.
The short cycle of large plant products is often “functional absorption”. Whether the office capacity of the Coze enterprise is integrated into the “Doubao job” or the small tools of Open AI are broken down into ChatGPT. From the perspective of product brands, they exit; but from the perspective of capacity, they continue to be replaced by a stronger and more core container. For large plants, this is a necessary cost for the reallocation of resources and a necessary option for maintaining organizational strength.
In large plant machines, no specific product is irreplaceable, and only the strategic objectives of the company cannot be wavered. When a product cannot independently sustain a billion-class imagination, or when its presence affects the advancement of the primary business, it is its inevitable fate that it is captured and cut off. The decisive cutting by byte of the Coze enterprise demonstrates the coldness and efficiency of the giant in its strategic focus.
Ecology of start-ups
The exit of the Coze corporate version has left a profound inspiration for the entire AI Agent track.
For a product team within a large plant, the long-term viability of a product is judged not only by the growth of external users but also by whether it is in the company's “principal”. Once overlapping or conflicting with the core strategy, the higher market share may also be recovered overnight.
For start-ups and small and medium-sized producers, the strategic convergence of large firms is a huge opportunity. The strategic convergence of large plants means that they abandon sub-markets that are “suffering and tired”, that require a strong alignment with industry-specific processes and that do not create platform-level monopolies.
For example, the Open-source AI platform Dify, which completed $30 million in financing in 2026, has chosen a life course that large plants are not willing to follow: handing over control to their clients and earning subscription fees for businesses to buy “data no-out” and “systems to govern”. Instead of competing for big individual developers, Dify went directly into large and medium-sized enterprises that were extremely sensitive to data security, allowing customers to load the system into their own rooms.
Avoiding the “main” route of the large plant, doing heavy, dirty work that the large plant is unwilling to do, serving long-tailed requirements that cannot be covered by standardized products of the large plant, could instead create a moat that the large plant cannot easily remove. Vertically and customarily, the best barrier to start-up companies is in the age of large factories pursuing “bold backbone”.
Agent's track is far from final. When AI Agent's pricing system moved from “software subscriptions” to “task delivery”, hundreds of billions of levels of market space were just opened.
The close-up of byte and the exit of the Coze corporate version are only the first sign of this long marathon. It reminds all practitioners that in the megagraph, products always serve strategy and may be pressed to stop at any time, while in the independent ring, products must serve their customers and only address specific pain points in order to have control over their survival. It is more important than blindly pursuing a larger and all-important position in the brutal AI phase-out.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.