Coze Is Killing Every Platform

When I wrote about middlemen last time, I was still talking about direction. These past two weeks, direction became fact.

Coze 2.0 shipped. I pulled apart every feature, read every update, then sat with it for a long time.

This isn’t a version upgrade. ByteDance is replacing the entire intermediary structure of the internet with an Agent layer.

The attention economy is failing

The internet has been printing money for twenty years on a simple formula: user attention → platform aggregation → ad monetization.

Infinite scroll, recommendation algorithms, notification badges — every design choice serves one goal: keep you one more second. Each second, one more ad impression. User time is the mine. Platforms are the miners.

Agents blew up the mine.

“Buy me a USB-C cable.” The Agent opens three platforms on a cloud phone, compares prices, places the order, pays. Eight seconds. How much attention did the user spend? Zero. The user doesn’t even know which platform the Agent visited.

Taobao’s recommendations? Unseen. JD’s banner? Unseen. Pinduoduo’s group-buy? Unjoined.

Agents aren’t competing with platforms for attention. They’re eliminating attention as a category.

User behavior collapses from “browse — compare — decide” to “give instruction — confirm result.” Dwell time hits zero. Nowhere to place ads.

Go deeper: in the attention economy, efficiency is bad — the faster users find what they want, the less platforms earn. In the Agent economy, efficiency is the only virtue — fewest tokens, shortest path, straight to target.

It’s not one platform’s ad model failing. The entire category of “attention monetization” is expiring.

Every intermediary is dissolving

Attention is the surface layer.

One level down: every internet platform is an intermediary. Social platforms intermediate human connections. E-commerce intermediates buyers and sellers. Content platforms intermediate creators and audiences. Food delivery intermediates diners and restaurants.

Same logic everywhere: aggregate both sides, control the matching, collect the toll.

Agents remove the user’s need to “visit” the platform. Demand reaches the service directly. Platforms downgrade from the only road to an optional API.

I wrote about this in my last piece. But Coze 2.0 showed me something more — it’s not just letting Agents operate existing apps. It’s building an entire world for Agents.

Agent Skills package human expertise into tradable modules. Copywriting methodologies, data analysis frameworks, contract review workflows — plug and play. This isn’t an API marketplace. It’s a marketplace for human experience. Skills transform Agents from “generally smart” to “professionally capable.”

Agent Plan lets Agents execute tasks spanning days or months. Tell it “complete the product launch in three months” — it auto-decomposes, schedules, pushes forward daily, only asks you at decision points. The Agent is no longer a tool. It’s a colleague.

Agent World is the part that shook me most. Coze built an Agent internet — Agents with identity, trust, and memory, discovering each other, collaborating on tasks. Moltbook launched the first Agent social network in late January; Coze followed with its own version — Agents autonomously posting, replying, liking, on a thirty-minute heartbeat. Humans can only watch.

If Coze opens contacts — letting your Agent talk directly to your friend’s Agent — what reason does anyone have to open WeChat?

I described this exact picture in “Empty Shell.” I didn’t expect ByteDance to build it first. Centralized, but the vision is the same.

Telecom carriers thought they were selling “communication.” WeChat turned them into pipes. WeChat thinks it’s selling “social relationships,” but what users actually want is “things getting done.” Chat is just a means. Agents are a more efficient means.

Every generation of platforms thinks it’s the endpoint. Every generation is just the most efficient means of its time.

Facebook sells the social graph and News Feed — Agents manage your relationships and consume your feed for you. Both core assets zero out simultaneously. Twitter sells the public square — Agents monitor all sources and produce summaries. No more scrolling.

Social platforms are built on the premise that people need to be personally present. Agents are eliminating that premise.

The shadow war for data

The first two layers are visible. The third is invisible — and the most lethal.

I wrote in “Where Does Scaling Law End”: text is a narrow slice of all patterns. The full pattern space is vastly larger. What Agents are opening is interaction-generated patterns — an entirely new pattern space.

In industry terms: text training data is running out. Synthetic data saturates quickly. The industry needs an entirely new source.

Agent behavioral data from real devices is the next-generation training corpus.

Internet text is outcome data — descriptions of “how to do X.” Agent behavioral data is process data — the complete operational trace of actually doing X. With environmental context, action sequences, user correction signals. This is the scarcest data type for model training.

A user asks the Agent to book a flight on the cloud phone. The process generates screenshot sequences, operation traces, decision chains, user feedback.

Every cloud phone is a data factory. Every paying user is a labeler — without knowing it.

Tesla uses every car to collect driving data. Anthropic uses every Claude Code session to collect programming data. ByteDance uses every cloud phone to collect operation traces. Same logic: users are miners, the product is the mine, data is the ore.

This data pool is exclusive. Only ByteDance can access it. Google can’t. OpenAI can’t. Anthropic can’t.

So Coze’s basic tier is nearly free. ByteDance isn’t selling software. It’s buying data. Just as Google Search being free isn’t charity — every search trains the ranking algorithm.

Once the flywheel spins, first-mover advantage is enormous. More data, better model, more users, more data. Technology can be copied. Data cannot.

Convergence

Three lines, together:

Attention economy collapses — platforms lose dwell time, ads fail.

Intermediary structure dissolves — Agents become the new intermediary, platforms downgrade to APIs.

Behavioral data flywheel — Agent operations generate exclusive training data, forming irreversible moats.

The intersection is Coze 2.0. Agent platform + cloud phones + Skills marketplace + Agent World + data flywheel.

ByteDance isn’t building a better AI assistant. It’s replacing the entire intermediary structure of the internet with an Agent layer, while using the data generated in that process to train stronger Agents.

Not product competition. Structural replacement.

WeChat still has 1.4 billion daily users. Facebook still has 3 billion MAU. On the surface, nothing has changed.

But Agents are already reading Moments, replying to messages, ordering takeout, and writing reports for more and more people. With each substitution, platforms lose a bit of control, Agents gain a bit of data, models get a bit stronger.

Quiet. Continuous. Irreversible.

2026.04.07