← Back to blog
AiAbout 5 min read

Coze Enterprise goes offline: a drag-and-drop canvas the models caught up with

Published Nov 8, 2026
Coze Enterprise goes offline: a drag-and-drop canvas the models caught up with

On October 29 the Coze (扣子) team published 《扣子企业版下线公告》, its sunset notice for Coze Enterprise. From November 30, Enterprise Standard (企业标准版) and Enterprise Flagship (企业旗舰版) stop selling to new customers; customers who signed but never ordered must talk to their account manager. It runs in two batches. Old plans bought on Volcano Engine, including the old Personal Pro, Enterprise and Team tiers, close on December 31, with unused balance refunded to the Volcano Engine account. New enterprise plans get until December 31, 2027, and no new purchase, renewal or extra seat may extend past that date.

One line in the notice is quiet and matters most: export and backup of existing conversations, memory, files, coding projects and low-code workflows "will be spelled out in a later announcement".

Products get shut down all the time. What is odd here is that the migration path is not written. For a company running dozens of bots and hundreds of workflows, the dates are only a reminder. The work sits inside that phrase, a later announcement.

Three years, three turns

Coze launched in February 2024, two months behind the overseas edition. It sold zero-code bot building: models, knowledge bases, plugins and memory split into nodes, with the user dragging lines to define each step's input and output. That was essential from late 2023 into mid 2024, because models could not break tasks down. Ask for "pull this company's earnings, work out gross margin, then email the sales director" and it dropped two of the steps. The canvas remembered them.

Coze then turned toward workplace AI, and later toward AI office work. On August 24 ByteDance folded TRAE and the Coze team into Doubao; TRAE Work, Coze and Doubao merged their workplace features, while TRAE IDE and CLI carry on as the coding line under the Doubao brand. The next day Doubao Work launched as the agent brand for productivity. Earlier, the Feishu product team had also moved into Doubao, and Feishu's built-in assistant Aily was renamed Doubao Work Companion.

So the shutdown, read after that August consolidation, is a closing move. Nobody was disbanded and nothing was thrown away. What got put away is the interface that made users assemble the flow themselves.

Same week, two exits and one demotion

Tencent was blunter. Yuanqi (元器) updated its notice on October 8 and stops service entirely on November 9, 2026. Launched in May 2024 at Tencent Cloud's generative AI summit, it sold drag-and-drop agents, plugins and knowledge bases with one-click distribution to WeChat and QQ. Half its pitch was the channel. The channel remains, the product does not.

Baidu took a third route. Miaoda's (秒哒) team was merged into Baidu Dazi (百度搭子) and Miaoda itself demoted to a Skill inside it. The name survives; a platform became a feature.

Three exits, one thing underneath: low-code orchestration is being eaten by the layer above it. Once a model holds long context, plans, calls tools and runs several steps, the request shifts from "wire up a flow for me" to "finish this for me". The first needs a canvas. The second needs a clear sentence.

19.3% of the market did not keep it

In IDC's June report, 《2025 中国智能体开发平台市场份额报告》, Volcano Engine (Coze) led with 19.3% revenue share, ahead of Baidu AI Cloud at 13.9% and Tencent Cloud at 12.6%. Together they hold under half, so the market was fragmented anyway. A leader walking away suggests the problem was not share.

When a front-runner pulls back on purpose, two readings are common: the market is smaller than it looks, or the value moved somewhere else. This looks like the second.

The canvas answered how to split a task when the model was not smart enough. By 2026 that answer got cheap. Hand the same task to an agent that plans, calls tools and retries, and its success rate is not necessarily lower than a hand-wired flow, at far less upkeep. The flow still exists, it just moved from the user's hands into the model. Approvals, audit trails and permissions now come straight from office suites, skipping the third party.

Low-code still has uses. What is depreciating is the skill of drawing lines. Knowing the business, and judging which step a human should block, still counts.

The real work is the move

Back to the later announcement. A platform that ran for two or three years holds more than expected: conversation logs, memory entries, uploaded files, coding projects, low-code workflows, plus the plugin keys and third-party authorizations hanging off those flows. Whether it exports, migrates and still runs on the other side decides whether a customer switches or quits.

A low-code workflow canvas with nodes and connecting lines, fading out as a chat window takes its place

Public information stops at "will be spelled out". For teams already on it, there is little to do, and it is concrete: export workflow nodes and prompts as text, keep local copies of key conversations and knowledge files, list every plugin authorization, then test whether those flows can be rebuilt inside a new office agent. Waiting for the notice leaves no room. December 31 is under two months away.

Platform consolidation has run several rounds in China's AI office track this year. Each round clarifies the product lines and means fewer apps installed. The cost is fewer choices, and someone always has to do the migration nobody wants.

The canvas was not wrong. It patched a hole that had already been patched.

Related articles