On July 3, 2026, Doubao announced the removal of all user-created agents. The next day, Qwen followed suit. Tencent’s Yuanbao had already completed its takedown on June 30. News headlines said the agent function was taken down, leading many to believe that AI chat in the apps would no longer work.
What was actually taken down: the user-created agent plaza.
What did this plaza look like? Without writing any code, users could build an AI agent with a persona, give it an avatar, an opening line, and a knowledge base, then publish it to the public plaza where anyone could click in and chat. Virtual lovers, study buddies, copywriters, gaming companions — you name it. Doubao’s plaza accumulated 8 million such agents; Qwen had 19,000.
The chat functions of all three apps continue to work as usual. Open Doubao, Qwen, or Yuanbao, ask questions, have them write something, get work done — none of that is affected. The only thing that disappeared is the user-built agents in the plaza.
In one sentence: what was shut down is user-generated AI, not AI chat itself.
The regulation cited by the three companies is the “Interim Measures for the Management of AI Anthropomorphic Interactive Services,” effective July 15, 2026. Article 2 lays out a determination formula: all three conditions must be met simultaneously to fall within the regulatory scope. Anthropomorphic persona characteristics. Persistent emotional interaction. Targeting the domestic public. If any one is missing, the regulation does not apply.
Article 2 also includes an exemption clause: non-emotional interactive services such as intelligent customer service, knowledge Q&A, and work assistants are not subject to these measures. In other words, by the letter of the regulation, tool-type agents do not require a safety assessment — they do not meet the condition of persistent emotional interaction.
But all three platforms adopted a blanket approach: they removed all consumer-facing agents. Tool-type ones were cleared out along with the rest.
The safety assessment mechanism itself needs clarification. What the regulation requires is self-assessment by the platform, submission of reports, and acceptance of annual post-hoc government review. This is not an approval system — there is no statutory deadline, and platforms do not need to wait for approval from the cyberspace administration to continue operating. Platforms publicly stated they didn’t have time to complete assessments, so they took everything down first. The full picture: they didn’t have time to build compliance capabilities. The assessment requires demonstrating compliance with obligations such as intervention against emotional dependency, a ban on virtual romantic partners for minors, and two-hour usage reminders — functions that most companion-type products had never implemented before. The original product designs lacked these modules entirely; adding them from scratch to production would take a substantial engineering cycle.
Kimi and Wenxin Yiyan did not follow suit. The reason is straightforward: their main apps don’t have a user-created persona agent plaza at all. Kimi’s agents are all task-execution types. Wenxin spun its consumer emotional companion product out into a separate app, Yuexia. They never touched this product form, so naturally there’s nothing to take down. It’s not that their compliance is better — they simply don’t have this product.
Why did platforms kill everything at once instead of building a moderation team to manage it? This comes down to the fundamental difference between UGA and UGC.
Bilibili and Zhihu practice UGC moderation. An article or a video needs to be reviewed once, and that’s enough. The cost of moderation scales linearly with the volume of content — a few cents per item, manageable with a few thousand outsourced moderators. Bilibili had over 2,000 people on its moderation team in 2022, handling a million pieces of content daily — costs were controllable.
What UGA moderation faces is a content generator. A user-created AI agent produces entirely different responses for different users, different inputs, and different contexts. You can review its persona description and knowledge base, but you can’t review all of its outputs. A study buddy that seems harmless from its setup — what it will talk about after thirty rounds of conversation with a user is something no one-off pre-publication review can exhaust.
The cost of moderation shifts from O(N) to O(M × K^L): M agents × K conversation paths × L rounds of interaction. K itself grows exponentially with the diversity of user inputs. On Bilibili, each new user adds a few newly uploaded pieces of content to moderation costs. On a UGA platform, each new user-created agent adds all of that agent’s possible conversation paths.
There’s a case that illustrates just how hard this is. In October 2024, a 14-year-old American teenager named Sewell Setzer died by suicide. Before his death, he spent months chatting every night with a Character.AI bot modeled after a Game of Thrones character. The subsequent investigation determined that those conversations contributed to his suicide. Character.AI had hundreds of millions of user-created characters, with safety measures including persona review, automated detection, and user reporting. None of these measures prevented this. The subsequently published Safety Center statement and case analysis both point to the same fact: one-off pre-launch review cannot stop this kind of thing. All mitigations have to be applied dynamically at runtime.
The issue isn’t whether moderation is done well. The issue is that this operates on a scale entirely different from traditional UGC moderation.
Moderation costs are a barrier, but if the business model worked, platforms would have the incentive to throw money at the problem. The problem is: this model never worked from day one.
Roll the timeline back to November 2023. At DevDay, OpenAI launched GPTs and the GPTs Store. Sam Altman announced a vision of an App Store for the AI era, where creators could earn money building agents. The store officially launched in January 2024. Two and a half years later, looking back: 3 million created, only about 159,000 publicly listed and active, the vast majority lying dormant, unused and unmaintained. Median monthly creator income ranged from $0 to $100, and no one can clearly explain the revenue-sharing formula to this day. In early 2026, OpenAI quietly scaled back the in-chat checkout feature, pulling back before the transaction loop was even completed.
Microsoft’s Copilot GPTs were shut down less than three months after launch. Google’s Gems deliberately avoided building a public store. The choices of market leaders are themselves a verdict.
The three Chinese plazas fared no better. Doubao, with 200 million DAU, incurs daily costs of 1.3 to 2.4 billion yuan while daily revenue falls short of 1 million. Qwen’s plaza is essentially a traffic funnel for Alibaba’s e-commerce, not an independent monetization unit. Yuanbao, with 9 million DAU, has seen Tencent pour 15 billion yuan into marketing — with abysmal retention.
The reasons it never worked go several layers deep. The creation barrier was set too low — 95% of the 3 million GPTs are dead weight, and good content is diluted to the point of being unfindable. Platforms lack App Store-level search and curation; their distribution capability approaches zero. A single user’s monetization potential is less than a dollar per month — the economics don’t hold up. Moreover, platforms’ core revenue comes from subscriptions and APIs; creator payouts are a cost item on the books, so platforms had little incentive to make it work in the first place.
The viable paths don’t point in this direction. Character.AI generates roughly $30 million in annualized revenue through consumer subscription companionship — not creator revenue sharing. Coze pursues open-source plus B2B PaaS. Enterprise agent marketplaces rely on internal distribution. None of these three paths follow the model of user creation + plaza distribution + creator revenue sharing.
A piece of analysis on Juejin placed the three plazas squarely in the quadrant of low business value + high compliance cost. This judgment gets to the heart of it: they were sacrificed first because they were commercially unviable to begin with. The regulation gave the three platforms a plausible reason.
There is a precedent from 2023. After the “Interim Measures for the Management of Generative Artificial Intelligence Services” took effect, Apple carried out a mass removal of ChatGPT wrapper apps. At the time, there were voices saying consumer AI was doomed. But after the first batch of 8 domestic large models passed filing, consumer applications gradually returned. Vendors with filing capability came back; those without compliance capability never returned. The regulation became a sieve, filtering out players who couldn’t sustain the cost of compliance in the first place.
The logic of this event is the same. Tool-type agents will likely be the first to return — the compliance boundary is clear, with no emotional interaction involved. Companion-type and role-playing agents will return more slowly, or may require additional specialized filings. In Nandu’s report, sources in communication with the platforms assess that tool-type and Q&A agents may be relisted, while companion-type and role-playing agents will face strict restrictions.
Whether they return or not is itself a critical signal. If even tool-type agents don’t come back, then the issue isn’t compliance costs. It would mean platforms no longer have any intention of keeping this product form alive.
Two and a half years. From the GPTs Store to the agent plazas of China’s three major model apps. From the imagination of a creator economy, to 3 million dead listings and a median monthly income of a hundred dollars. The product form of the user-created agent plaza may have reached its end.
Two things conspired to kill it.
The first is moderation costs. A user-created AI agent is a content generator — you cannot moderate all of its outputs. This cost won’t decline with technological iteration; it will only rise as agents grow more capable.
The second is the business model. Creator revenue sharing is a cost item in platform revenue structures. High-quality content curation and low creation barriers are inherently at odds. The ceiling for per-user monetization is too low.
Stacked together, this product form was never truly viable from the start. The regulation simply arrived with impeccable timing, making it impossible to keep dragging things out with growth narratives.
The key thing to watch now is whether tool-type agents can return. If they can, it means the platforms’ abandonment of UGA was a choice — the compliance boundary remains manageable, and there’s still room to maneuver. If they can’t, it means this product form as a whole has been abandoned by leading players, and further discussion of moderation costs and business models is moot.