Coze’s Enterprise Shutdown Explained: From Origins to Now

Why it is suddenly everywhere

1. 01

   A Zhihu question asking whether Coze’s enterprise shutdown means AI workflows have lost their market pushed ByteDance’s October 29 notice back into the discussion in early November.

2. 02

   The official notice has two separate tracks: legacy Volcano Engine plans end on December 31, 2026, while the newer Enterprise Standard and Flagship plans last until December 31, 2027. New enterprise sales pause on November 30, 2026.

3. 03

   This is not an announcement that every Coze product is closing. The notice covers enterprise plans; the separate Coze Studio repository still offers visual agents, plugins and workflows under Apache 2.0.

4. 04

   Tech Planet cites IDC figures giving Coze a 19.3% lead in China’s 2025 public-cloud agent-development platform market. Yet IDC put that whole public-cloud category at roughly RMB 250 million, against RMB 1.75 billion for private deployments.

The timeline

1. 2023 H2Bot Studio starts inside ByteDance

   GeekPark’s reconstruction says the team first imagined AI-assisted programming, then chose a more feasible no-code chatbot builder. Coze appears overseas in late 2023.

2. 2024-02Coze launches in China

   The Chinese product lets people create bots with prompts, plugins and little or no code. Early creation is easy; retaining useful bots proves harder.

3. 2024–2025Workflows find the durable users

   Team reporting describes rising, stable workflow use in data processing, customer service, documents and internal coordination. The product shifts toward “serious developers” solving work problems.

4. 2025-04Coze Space tests the office-agent turn

   An April 18 beta combines a general agent with specialist templates and Lark integration. KrASIA reports a usage spike beyond allocated compute within hours.

5. 2025 H2Enterprise plans arrive

   Tech Planet dates the enterprise edition to the second half of 2025. Standard and Flagship packages turn the work platform into a standalone organizational SaaS offer.

6. 2025-07-29The engine goes open source

   ByteDance releases Coze Studio and Coze Loop under Apache 2.0, covering visual agent development, workflows, evaluation and observability.

7. 2026-01Coze 2.0 reframes the product

   Skills, long-running Agent Plans and Coze Coding move the pitch from assembling bots to workplace AI and natural-language development.

8. 2026-08-24The team joins Doubao

   ByteDance confirms that TRAE and Coze teams are joining Doubao. Workplace capabilities will be integrated; TRAE IDE and CLI continue under Doubao’s programming line.

9. 2026-10-29The enterprise wind-down becomes official

   Coze publishes its phased timetable, citing enterprise-product integration and plan upgrades. Replacement products and resource-export arrangements are left for later notices.

The backstory

The first hit was not the bot store

Coze began with an appealing promise: describe a bot, add some tools, and publish without learning a software stack. GeekPark’s history says the internal Bot Studio team had originally wanted something closer to today’s AI coding products, but judged the 2023 models and engineering scaffolding inadequate. A prompt-driven bot builder was the practical starting point.

Creation turned out to be the easy part. A person could make a bot in minutes, feel the small thrill of having “built software,” and never return. GeekPark describes weak retention among the many user-generated bots. The more revealing signal sat elsewhere: workflow calls kept rising, and their users stayed. Data processing, customer-service automation, document handling and coordination were less glamorous than a bot store, but they recurred every working day.

That explains a seemingly odd design choice. Workflows make an AI system look more like conventional software: a person arranges nodes, decides the order and limits what the model can do. The team had initially worried that this was not “AI native.” In practice, those constraints were the point. Early models were unreliable at long, multi-step tasks; a visible flow supplied guardrails and made failures easier to locate.

Coze Space then pulled the product in the other direction. Its April 2025 beta let a general agent plan tasks, call tools and work with specialist templates. KrASIA’s adapted LatePost report describes both the ambition and the rough edges: the agent could build reports and prototypes, but testing also produced outdated answers and hallucinations. Coze’s history is therefore not a straight replacement of workflows by agents. It is a repeated adjustment of how much planning belongs to the model and how much remains in software.

A market lead with a small denominator

The shutdown looks paradoxical only if “number one” is allowed to float above the size of the market. Tech Planet, citing IDC’s 2025 market-share report, gives Coze 19.3% of China’s public-cloud agent-development platform market, the largest vendor share. IDC’s reported category revenue is much more sobering: approximately RMB 250 million for public-cloud platforms, still in the pilot and early-commercialization stage.

Private deployment was a different business. IDC put that category at RMB 1.75 billion in 2025. Its buyers were looking beyond an easy builder toward testing before release, monitoring afterwards, permissions, content safety, token-cost governance and management of agents from different sources. Those requirements favor deep enterprise delivery, not just a polished canvas. They also make a simple comparison between a hosted subscription and a private platform misleading: the implementation and revenue structures differ.

This is the commercial backdrop, not proof of a single cause. Coze’s official notice says “product integration and plan upgrades,” not “the market was too small.” But the figures explain why a leading share did not guarantee an independent enterprise suite. The same AI economy was also learning to sell usage in new packages, as the livestreamed AI token phenomenon showed; turning that activity into durable enterprise software revenue was a separate problem.

The organization chart moved before the shutdown notice

The clearest sequence is organizational. On August 24, ByteDance confirmed that the TRAE and Coze teams would join Doubao. Coze’s workplace capabilities were to be integrated with Doubao, while TRAE’s IDE and command-line tools continued as a programming product line. The company told National Business Daily that the goal was to coordinate product and technical resources and that existing user rights would not be affected.

Two months later, the enterprise-plan notice supplied the product-level consequence. Jiemian reads the episode as ByteDance replacing parallel AI experiments with a “thick main trunk”: Doubao for the primary assistant and work entrance, Volcano Engine for infrastructure and enterprise delivery. That is an interpretation, not wording from Coze’s announcement, but it fits the order of events better than the idea that a workflow feature suddenly stopped working.

It also has a consumer-internet parallel. The ByteDance Time debate was about attention collecting around a few major apps. The Coze story is about work collecting around a few AI entry points. A separate enterprise subscription can survive as a feature or capability inside that arrangement without surviving as an independent package.

What ends, and what does not

The official boundary matters. Enterprise Standard and Flagship stop accepting new customers on November 30, 2026 and fully retire on December 31, 2027. Older personal-advanced, enterprise and team packages bought through Volcano Engine retire earlier, on December 31, 2026. Existing customers can continue within their contracts until the relevant date; renewals and added seats cannot extend beyond the final deadline.

The announcement does not declare the closure of the entire Coze website, personal plans or the open-source projects. Coze Studio’s public repository still contains the visual agent builder, workflow engine, plugins, knowledge bases and APIs under Apache 2.0. ByteDance also open-sourced Coze Loop for evaluation and observability in July 2025. Neither fact guarantees permanent maintenance or compatibility with a future hosted product; it simply means “Coze is shutting down” is broader than the evidence.

Nor does the episode prove that workflows are obsolete. Better models can generate plans and even draw the workflow themselves, reducing the value of dragging every node by hand. Production systems still need repeatable steps, permissions, logs and hard limits. The division of labor examined in Jev’s bounded-decision model makes the same point: generative models can plan, while code and specialized components still enforce branches, calculations and constraints. The workflow is moving from a hand-drawn diagram toward generated, observable infrastructure.

One practical uncertainty remains. As of November 8, Coze says arrangements for exporting and backing up conversations, memory, files, coding projects and low-code workflows will be announced later. The dates are clear; the migration path is not yet complete.

Key facts

Notice published

Oct 29, 2026

New enterprise sales pause

Nov 30, 2026

Legacy plans end

Dec 31, 2026

Standard & Flagship end

Dec 31, 2027

Coze public-cloud share

19.3% · IDC via Tech Planet

Public-cloud market

~RMB 250M · 2025

Private-deployment market

RMB 1.75B · 2025

Open-source license

Apache 2.0

What people are saying

Common questions

Is all of Coze shutting down?

No such full-platform closure is stated in the October 29 notice. It names legacy hosted plans and the newer Enterprise Standard and Flagship plans. The personal product and the Apache-licensed Coze Studio and Coze Loop projects are not listed as part of this wind-down, although their future product arrangements still require separate confirmation.

Can existing enterprise customers keep using the service?

Yes, within their contracts and before the relevant shutdown date. Legacy Volcano Engine plans end on December 31, 2026; newer enterprise plans end on December 31, 2027. Renewals and added seats cannot run beyond that final date. Unused legacy-plan service beyond the 2026 deadline will be refunded automatically to the Volcano Engine account.

Does this prove AI workflows have no market?

No. It puts pressure on one business model: a standalone hosted suite built around manually configured agents. IDC still describes demand for private deployment, workflow generation, evaluation, observability and governance. Models may draw more of the flow themselves, but predictable business processes and hard controls remain valuable.

Can users export their bots and workflows now?

The notice does not yet provide a complete answer. It says export and backup arrangements for conversations, memory, files, coding projects and low-code workflows will be specified in a later announcement. As of November 8, that migration detail is still pending.