@shellywangcat: Over the past couple of months, several pieces of news landed all at once. ByteDance's Coze Enterprise Edition stopped accepting new purchases and is being phased out, with the team folded into Doubao; Tencent's Yuanqi shut down on November 9 — a bit over two years from launch to closing; Baidu's Wenxin Agent Platform shut down part of its creation entry points. Seeing the news about Coze, I felt quite a pang. We built something very similar, and at the end of last year...

Over the past couple of months, several pieces of news have piled up.

ByteDance’s Coze Enterprise Edition has stopped accepting new purchases and will be phased out, with the team merged into Doubao. Tencent’s Yuanqi shut down on November 9. From launch to closure, just over two years. Baidu’s Wenxin Agent Platform has turned off part of its creation entry point.

When I saw the news about Coze, I felt a wave of sentiment. We had built something very similar, and at the end of last year, we too pressed the pause button.

Here are some of the things I’ve been thinking about over these two years.

1. What made Coze obsolete

Many people say workflow has been replaced by AI. But I believe workflow never went away. What got eliminated was “humans drawing workflows.”

Two years ago, models couldn’t plan. You’d ask them to do a five-step task, and by step three they’d veer off course. So platforms gave you a canvas and let you draw out every step yourself. Drag a node, connect a line.

Looking back now, all that dragging and dropping was paying for the model’s clumsiness. It was a transitional form carried over from low-code. Once models could plan, the diagrams you drew on the canvas became plans worse than what the model would produce.

Coze’s own fate is the answer: first it pivoted to AI coding — speak a sentence, and an Agent writes the code; then it merged into Doubao Work. The craft everyone used to learn, “how to build a workflow,” has now been absorbed by Agents too.

2. Three kinds of canvases, three different fates

But not all canvases are dead. Categorized by what gets drawn on the canvas, there are three types.

The first type: drawing reasoning. Each node is a prompt; the connections mean “think about this first, then think about that.” Both Coze and Yuanqi fall into this category. This type is fading away.

The second type: drawing connections. Nodes are APIs and systems; the connections represent where data comes from and where it goes. “When an email attachment arrives, save it to cloud storage” — that kind of thing doesn’t need a model to think. It needs cheap, reliable execution. Zapier, n8n, Dify live in this category. This type is still alive. Though n8n is valued at $5.2 billion on disclosed annual revenue of roughly $40 million — and a fair chunk of that valuation is story.

The third type: don’t draw anything, just write code. But underneath there’s a runtime layer that manages state, retries, and where to pick up from after a crash. Temporal, Restate, Inngest. This layer is the hardest. Temporal’s valuation reached $12.55 billion this year, with annualized revenue exceeding $250 million and growth over 200%. OpenAI and Netflix both use it.

The pattern is simple:

• The closer you are to “drawing reasoning,” the faster you die.
• The closer you are to “guaranteeing execution,” the better you survive.

And I’ve since used one question to check everything I build: if the next generation of models is ten times stronger, does my business get better, or disappear?

Canvases for drawing reasoning: disappear — because the models can think for themselves. Runtimes: get better — the longer an Agent runs and the more tools it calls, the more places there are for things to go wrong, and the more you need something to catch it.

3. The “Coze” we once built

What we originally set out to do: business users describe a requirement in one sentence, and the system turns it into a process that executes reliably. No learning prompts, no dragging nodes.

Technically, we took many detours. At first we had the model write code at runtime — unstable. Later we switched to generating the flow during the planning stage, then assembling pre-validated skills. Flows that proved out were hardened into stable code.

In August this year, ByteDance released Doubao Work. I went through it carefully. Self-decomposition of goals + a skill store + connectors. Common flows get solidified and called directly next time.

The skeleton is almost identical to what we built back then.

So the idea wasn’t wrong. Where did we fall short?

The moment Doubao Work logs into Feishu, three years of enterprise documents, chat history, and meeting notes are all right there. The permission system comes ready-made. The execution environment includes cloud computers — close the lid and the task keeps running. The entry traffic is massive and free.

Every customer we onboarded, we had to first convince them to connect their spreadsheets, ERP, and backend data one by one.

Now look at the three giants:

• Tencent’s WorkBuddy connects WeChat and WeCom.
• Alibaba merged three products into Qwen Office, also integrating with enterprise IM.
• ByteDance’s Doubao Work connects to Feishu.

They’re all doing the same thing: hanging Agents on their own data. The Agent follows the data. Wherever the data lives, that’s where the Agent’s home is.

MaybeAI was the right idea. But it had no home.

4. The trend we chased

When Gemini 3.0 came out at the end of last year, its image capabilities jumped dramatically. When we had to pivot, we spent some time on an AI design tool for cross-border e-commerce: swap models, swap scenes, generate images according to platform rules and localized styles.

Later we saw that Alibaba’s own e-commerce image tool shut down too. Huiwa posted a service termination notice; Luban went onto the reduction list.

Alibaba isn’t short on e-commerce data, traffic, or people who understand e-commerce. It shut down anyway.

The reason is plain: e-commerce image generation is becoming a default model capability. Sellers open a general-purpose model and they’re done — they don’t need to enter a specialized tool anymore.

Building tools around the latest model capabilities eats into a time gap: the model just learned it, the product hasn’t caught up. The next version ships, the gap disappears — roughly six months. The newer the capability, the sooner it gets eaten.

Many people say business know-how is worth more than everything. I now think that’s only half right.

The know-how you can write down — prompt templates, category rules, platform compliance — will eventually be fed to the models anyway.

What’s left is the stuff you can’t write down.

5. When tools die, where do the people go?

As tools get eaten layer by layer, what we’re seeing is that value flows more and more to people.

The hottest role this year is FDE: sit on-site with the customer, find requirements, connect systems, get the Agent running.

LinkedIn statistics show this role grew 42x from 2023 to 2025. ByteDance is offering up to 1.05 million RMB in total compensation for Doubao FDEs. Anthropic says it plans to train 10,000 FDEs before the end of 2027. In late August, the Ministry of Industry and Information Technology issued a document that, for the first time, wrote FDE into policy.

The cross-border e-commerce world is the same. Most of the bosses I’ve been talking to lately don’t want to buy platforms anymore. They want to hire one person in-house whose sole job is exploring and hardening Agents. Some call them an agent engineer.

But there’s a two-sided coin.

The FDE model was invented by Palantir. But few mention what came next: once the platform matured, standardized deployments replaced a great deal of customization, gross margins climbed above 80%, and it actually reduced its dependence on FDEs.

FDE is the dirt-road stage. It only counts once the dirt road becomes a highway.

Many FDE roles in China are just implementation or pre-sales with a new name.

Matchmaking platforms are the same. One end takes enterprise demand, the other dispatches service workers. If all they charge is a referral fee, next time the customer will skip the platform, experience stays with individuals, and nobody pays when delivery goes wrong. Those that survive need to be like a general contractor in construction: sign off on the result and take responsibility when things go wrong.

6. The three things left over

So what, exactly, won’t get eaten?

We break it into three questions:

1. If this step goes wrong, how big is the cost?
2. If it goes wrong, how do we know?
3. When something breaks, who do we hold accountable?

The first is about how much risk you dare to take. “Send an email to a customer” — a bank needs approval, a startup just hits send. Same task, the answer depends on whose business it is.

The second is what counts as the source of truth. Whether the reconciliation balances depends on the bank statement. Whether an image is compliant depends on platform review. The truth lives outside the system, and the model can’t be its own judge.

The third is who signs off.

Models will increasingly know how to be stable. How stable to be and what counts as truth will always stay in human hands.

7. The boss can build it themselves now

There’s another shift, even faster than FDE.

Among the cross-border bosses I talk to, more and more are opening Codex themselves to build flows: log into the backend, download reports, clean the data, format it into the table they want — done in a few days.

Platforms, image tools, and even some of the agent engineer’s job have been eaten by “boss + coding Agent.”

But I’ve also seen the other side.

The script lives on the boss’s laptop. Lid closed, and tomorrow’s report never arrives.

The platform renames a field, the script doesn’t throw an error — one column just goes blank.

Worst of all: the numbers are wrong but the report still looks perfectly tidy. Thirty SKUs written three different ways never got merged, inventory is undercounted, and nobody notices.

Building it is nearly free. Keeping it running reliably is not.

8. We decided to become the water pipe

So we changed our stance.

Not the home. Customer data stays in their WPS and Feishu spreadsheets — that’s where they work every day, and we’re not going to grab it. Just the water pipe: delivering outside data into their spreadsheets cleanly, on time, and reconciled.

This matches what many people now want from AI hardware: hardware that honestly collects data and exposes an MCP interface for the Agent to use — far better than stuffing a half-baked Agent inside the hardware.

We too fell into the classic developer trap of always wanting to be the platform, to be the entry point.

But a water pipe has its own trap: a pipe anyone can connect to gets its price squeezed. It needs to have something others can’t replicate.

Once we figured that out, we decided to do one very small thing: help small and micro-business finance teams with reconciliation and profit.

Two years ago, we thought we were building a platform.

Later we found out platforms get eaten by models, tools get eaten by models, and even building workflows — bosses can now do that themselves.

What won’t get eaten: the dirty, grueling, directly-money-related places where someone has to take responsibility when things go wrong.

We plan to stick around there.

Cangshu (@Sither_V): Coze is going offline, and I literally worked on Coze at ByteDance. My profile page still says “Coze Workflow Engine Developer.” Next year when I post that, nobody will even know what Coze is.

I’m feeling sentimental about this thing coming to an end. Let me share a few things you won’t see from the outside:

1. workflow wasn’t replaced by AI — it was replaced by “the model can plan for itself”