Doubao, Qwen and Yuanbao All Shut Down Their AI Agents — From July 15, Your AI Companion Is Gone

On the evening of July 4, many people opening the Doubao and Qwen apps received a calmly worded notice: agent features will be officially taken offline on July 15. That came just five days after Tencent's Yuanbao shut down its agents on June 30. The three giants hit the same brake at the same time, all pointing to the same day: July 15, 2026, when China's 'Interim Measures for the Administration of Anthropomorphic Interactive Services' takes effect. This is not a product iteration — it's an industry-wide, collective netting operation.

💡 What You Will Learn

On the evening of July 4, many people opening the Doubao and Qwen apps received a calmly worded notice: agent features will be officially taken offline on July 15. That came just five days after Tence

What's happening isn't just shutting a feature — it's clearing data. Doubao's notice is restrained ('due to product feature adjustments'), but the behind-the-scenes action is anything but gentle. From July 15, Doubao fully closes agent creation, editing, and online chat entry points, and all user-created agents stop operating. ByteDance gives users a 3-month buffer: from July 15 to October 15 you can export your data; after that it is permanently deleted and unrecoverable. Qwen is quicker — anthropomorphic-interaction agents and user-created agents went offline as early as July 10, with a full shutdown on July 15, after which users can no longer access any agent configuration or chat history. Yuanbao moved earliest, quietly pulling its agents on June 30.

The new regulation: on April 10, China's Cyberspace Administration (CAC) together with the NDRC, MIIT, public security, and market regulators issued the 'Interim Measures for the Administration of Anthropomorphic Interactive Services,' effective July 15. It targets AI that 'simulates a natural person's personality traits, thinking patterns, and communication style to provide continuous emotional interaction services.' In plain terms: AI pretending to be a person, chatting with you, cheering you up, and making you dependent — that's what's being regulated. The red lines are clear: no virtual relatives, virtual companions, or other virtual intimate-relationship services for minors; minors under 14 using anthropomorphic AI must have parental or guardian consent; platforms must build a minor mode with usage-time limits and periodic reality reminders; and services with 1 million registered users or 100,000 monthly actives must pass a security assessment first.

Behind these clauses is the mess left by over a year of runaway growth in AI emotional-companion products. On June 26, the Shanghai Cyberspace Administration published the first-phase results of its 'Qinglang' campaign against AI app chaos: more than 14,000 violative agents were removed — equivalent to wiping out the AI content pool of a mid-sized app store. MiniMax was one of the key targets, with its 'one-click undress' and sports-betting agents continuously cleaned up. And these are only the verifiable numbers; agents masquerading as 'virtual lovers' or 'emotional mentors' while inducing minors toward extreme content or harvesting private data far outnumber what was published. The logic is actually simple: once anthropomorphic AI products scale, they stop being a 'niche emotional experiment' and become 'social infrastructure for hundreds of millions.' Infrastructure can't go without safety guardrails.

But not all AI is cut with one knife. A key exemption clause reads: services providing smart customer service, knowledge Q&A, work assistants, learning/education, and research that do not involve continuous emotional interaction are not covered. In other words: kill the 'chatting', keep the 'work'. This distinction is critical. When asked whether tool-type agents would be relisted later, both Qwen and Yuanbao declined to respond — and silence is itself a response. The compliance cost is too high: with mass user-created agents whose content is wildly varied, platforms would need full-chain real-time review, role filing, and risk tracing, which is nearly unworkable at current technical and operational levels.