Before large language models arrived, white-collar workers understood office automation in a very visual way. People were used to looking at flowcharts and clicking software buttons step by step with a mouse. When large models first went mainstream two years ago, product managers followed that same old habit. Since the goal was to have AI do the work for humans, they built a web canvas where people could draw their workflows as flowcharts with a mouse and let the AI step through the boxes. Coze, the agent-building platform under ByteDance, took off on the back of this canvas. In the two and a half years since its launch, users have built over 100 million agents, bringing together 3 million developers and more than 700 plugins, capturing the No. 1 share of 19.3% in China’s public cloud agent platform market.
Yet lively usage numbers failed to translate into real revenue. On October 29 this year, ByteDance posted the Coze Enterprise discontinuation announcement on the Volcano Engine official website, announcing that current enterprise plans will stop new sales on November 30 and shut down completely by the end of 2027. Opening the books disclosed by market research firm IDC makes the reason clear. Last year, China’s entire public cloud agent platform software market totaled just 250 million RMB; even in first place, Coze brought in only about 48 million RMB in annual software revenue.
Some might point to overseas peers as a counterexample: n8n, a German tool that also connects boxes with lines on a web page, just secured strategic investment from European software giant SAP this May, doubling its valuation to 5.2 billion USD; investors revealed that its annual recurring revenue exceeds 100 million euros, or roughly 780 million RMB—16 times Coze’s public cloud software revenue. Doesn’t that prove node-and-wire canvases still have a bright future? Quite the opposite. Although n8n wears the same node-connecting canvas face as Coze on the surface, underneath it is a completely different product category: n8n sells deterministic engineering plumbing that moves data across enterprise systems, whereas Coze sells prompt flowcharts where humans use a mouse to teach AI how to think step by step. It is precisely this latter category—a transitional product stuck between old human habits and new AI capabilities—that has reached the end of the road in just two and a half years.
Many people have tried Coze firsthand, found dragging a few boxes around convenient, and wondered why enterprises refuse to adopt it. Suppose you work in operations at an e-commerce company and want AI to automatically check customer refund emails every day, look up internal rules, draft replies, and post them to a Feishu group. On the afternoon of Day 1, you drag 4 boxes onto Coze’s web interface and connect them in a line: receive message, query refund documentation, call a large model to draft a reply, and send to Feishu. In just 15 minutes you have automated the workflow, and your whole team applauds.
By the second week, however, your boss hands you two real business rules: if the order query API times out, wait 5 seconds and retry 3 times; if a customer has requested a refund 2 times within 3 months, route the case to manual review. On a canvas where you cannot directly write loops and exception-handling code, what started as a 4-box flowchart instantly balloons into a tangled spiderweb of 50 or 60 boxes. What hurts enterprises even more is Coze’s official enterprise billing documentation: the Enterprise Standard plan starts at 980 RMB per month, the Flagship plan starts at 8,980 RMB per month, and every execution deducts additional credits. Everyday conversations cost dozens of credits, calling a cloud phone costs 200 credits, and generating a video costs 6,000 credits. Even if a workflow fails at the 58th box due to a network hiccup, all the credits burned by the preceding steps are still deducted in full.
An even bigger landmine goes off in the third week. The prompts and extraction rules inside every box on the canvas were painstakingly tuned to the quirks of a specific large model. Yet Coze’s underlying models depend directly on Volcengine Ark, ByteDance’s large-model compute platform. According to Coze’s official model deprecation notices, between March and October this year alone, Coze retired more than 30 mainstream models in 5 batches as Volcengine Ark shut down older models. On August 31 alone, the platform turned off over 10 versions, including flagship Doubao releases as well as multiple models from Kimi and Zhipu; on October 22, it retired another 2 lightweight Doubao models. The official documentation states plainly: once a model is taken offline, live agents, workflows, and applications using that model will no longer run properly. Every model upgrade turns into a production outage, forcing administrators to manually swap models in the backend and retest everything.
By Day 30, the platform announces that the enterprise edition is shutting down. When the team tries to export and back up the canvases and customer memories they spent half a year wiring together late into the night, a line in the announcement FAQ leaves them stunned: official support states that export and backup arrangements for existing conversations, memories, files, coding projects, and low-code workflows will be specified in a subsequent notice. At the very moment discontinuation was announced, there was no ready-made bulk export tool. Meanwhile, the billing documentation explicitly states that 15 days after a subscription expires or is canceled, the system automatically destroys enterprise memory stores with no possibility of recovery. While the personal and open-source editions remain in operation, companies that paid for the public cloud enterprise edition cannot even get a bulk export tool. Built around the human habit of clicking a mouse, web flowchart canvases buckle under complex logic and model turnover as soon as they enter real production environments, while locking core assets inside someone else’s database.
Looking only at web screenshots, many people lump the German tool n8n, the American tool Zapier, and ByteDance’s Coze together as direct peers. Strip away the surface, however, and you find that they have belonged to two completely different product categories since day one: Coze uses flowcharts to hard-wire the reasoning steps of large models, whereas n8n sells tightly bolted enterprise data plumbing and never bet its future on a chatbot canvas.
When founder Jan Oberhauser wrote the first line of n8n code in late 2018, today’s large-model boom did not yet exist. The problem he set out to solve was the unglamorous work of moving data between enterprise systems: how to reliably connect customer emails, SAP finance systems, databases, and chat apps; how to automatically refresh login credentials when they expire; and how to resume execution after a power outage and reboot without double-charging customers. In an early growth retrospective, he specifically highlighted the 80% trap of no-code tools: ordinary canvases make building the first 80% feel effortless, but hit a wall on the final stretch required for real production, forcing you to spend 5 times as long paying down the technical debt.
To solve that problem, n8n took the exact opposite approach from Coze. Its web canvas is merely a thin interface. You can select all the boxes on the web page, press copy, paste them into a text editor, and get a plain-text code file that can be checked directly into your company’s code repository and run for free on internal servers. Whenever the visual boxes fall short, you can drop in a code node at any time and write JavaScript or Python scripts to handle complex logic.
In its business model, n8n remains model-agnostic. Enterprises simply plug in their own model API keys, and the platform charges only for the number of times an entire business pipeline finishes running end to end—or collects on-premise software licensing fees from more than 1,400 large enterprises. Backed by over 400 enterprise software connectors, it supports 100 million euros in annual recurring revenue and 1.7 million monthly active builders with a team of roughly 185 employees. Its American peer Zapier, armed with over 9,000 software connectors, has stayed profitable for 12 consecutive years after raising just 1.3 million USD, holding annual recurring revenue steady above 310 million USD.
Back in China, data from research firm IDC shows that 87.5% of Chinese enterprise spending on agent platform software goes toward private deployments installed inside companies’ own server rooms. Putting a web canvas on the public cloud for employees to draw flowcharts can neither enter large corporate intranets nor handle complex system integrations, so it naturally fails to capture enterprise budgets.
If the only missing pieces were enterprise connectors and code export, ByteDance could easily hire a few engineers to add them to Coze. Why did ByteDance skip that path and instead shut down Coze Enterprise outright, ceasing to sell it as a standalone product line?
Because two years ago, AI was like a forgetful intern who didn’t know how to use tools on its own. If you gave it a slightly complex task, it would immediately forget the rules. Humans didn’t dare let it make decisions by itself, so they had to use a mouse on a web page to draw rigid boxes—step one search for information, step two extract a table, step three branch by amount—telling the model exactly what to do at every step.
This transitional layer wedged between human habits and AI capabilities lost its reason to exist this year. Today’s large models have gained two game-changing capabilities: planning task steps on their own, and writing code directly to operate a computer. Whether it is overseas command-line coding agents like Claude Code and Codex, or ByteDance’s own coding tool TRAE, they can write scripts, call APIs, and debug their own errors directly inside a controlled computer sandbox. Take the weekly task of reconciling 50 supplier price quotes in different formats and flagging abnormal price hikes: previously on the Coze canvas, you had to drag out 30 to 40 boxes to handle text splitting, loops, and format conversions, and if a single supplier changed its column headers, the entire canvas would throw an error and grind to a halt. Today, you simply write a dozen lines of plain-language instructions, and the agent immediately writes a 30-line Python script inside a virtual computer sandbox, runs the loops, adapts to header changes, checks errors, re-runs on its own, and hands you the reconciled spreadsheet 2 minutes later.
A 30-line snippet of dynamic code is far more flexible than 60 rigid connected boxes on a web page. Once you understand this math, ByteDance’s move in late August this year makes complete sense: ByteDance officially merged both the Coze team and its coding tool TRAE into the Doubao organization, launching a unified workplace entry point called Doubao Work, while handing large government and enterprise clients that require on-premise installation over to Volcano Engine’s private deployment team.
Inside Doubao Work, the front-stage lead is now TRAE, a coding agent that can write code and run commands. Coze has stepped backstage, unbundling the cloud computer sandboxes, cloud phone sandboxes, and over 700 plugins it accumulated over 3 years into a tool library that the coding agent can call on demand. In mid-September, the overseas edition of Coze went so far as to axe 9 core workflow features, including scheduled triggers and long-term memory. Zapier in the United States is making the exact same pivot, opening up its library of over 9,000 software connectors so Claude and ChatGPT can call them directly as ready-made hands and feet. Using a mouse to draw flowcharts for AI was only a temporary crutch; once the crutch is thrown away, complex judgment belongs to code written on the fly by AI, and fixed actions belong to underlying software interfaces.
Once we understand this technological shift, the most direct takeaway for our daily work or AI product development tomorrow is how we accumulate assets. Instead of spending time dragging and wiring boxes on someone else’s web canvas, we should keep two genuinely portable assets in our own hands.
First, write down your daily working rules and domain know-how as plain-text manuals saved on your own computer. Stop locking your prompts, judgment criteria, and business workflows inside any big tech company’s web boxes while waiting around for a future export announcement. Create a folder on your own machine, use ordinary plain-text Markdown files to spell out your task goals, boundary rules, and acceptance criteria, and track them in a code repository. Those plain-text assets remain under your personal control. If Doubao works best today, feed them to Doubao; if Claude or DeepSeek releases a better version tomorrow, you can switch over in a second and keep running.
Second, package fixed actions that interact with external systems into standalone small scripts or standard tools. What coding agents do best is reading your manual and writing code on the spot to string a workflow together, but they shouldn’t be poking directly at your production database, nor should they have to rediscover how to log into your company email from scratch every single time. Wrap fixed actions like querying databases, sending Feishu notifications, and scraping web pages into clean command-line utilities with well-defined inputs and outputs, or use the open-source Coze kernel and self-hosted n8n to set up fixed plumbing on your own servers. Equip AI with secure computer sandboxes and reliable hands and feet, and let it write the code to run the rest of the workflow on its own.