China — AI Companion Regulation Forces ByteDance and Alibaba to Kill Agent Features
AI relevance: China's new anthropomorphic AI regulation demonstrates how compliance requirements can be architecturally incompatible with persistent-memory agent design — a global precedent for agent governance that directly shapes how AI agent platforms handle user data, identity, and autonomy.
What happened
- On July 6, 2026, both ByteDance's Doubao and Alibaba's Qwen announced their AI agent creation features will be discontinued on July 15, 2026 — nine days from now.
- The shutdowns coincide with China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services, co-issued in April 2026 by the Cyberspace Administration of China and four partner agencies.
- The regulation targets services that "simulate the personality traits, thinking patterns, and communication styles of real people to provide continuous emotional interaction" — standard productivity chatbots and customer service bots are explicitly excluded.
- Doubao's agents allowed users to create distinct AI identities with separate memory stores, communication styles, and expertise areas — precisely the persistent-memory, emotionally-continuous architecture the regulation was written to govern.
- The compliance requirements include: mandatory anti-addiction pop-ups after 2 continuous hours of interaction, instant-exit mechanisms the platform must honor immediately, and real-time detection of over-dependence signs with on-screen reminders that the service is artificial.
- These requirements are architecturally at odds with persistent-memory agent design — a platform cannot simultaneously build the feature that creates emotional attachment and install the mandatory friction designed to interrupt it.
- ByteDance judged rebuilding under a new compliance architecture more practical than patching the existing agent system. Alibaba made the same calculation for Qwen.
- After July 15, users cannot create new agents and existing agents stop functioning. Doubao offers read-only access to configurations/chat histories until October 15, 2026. Qwen announced no grace period — all agent data will be permanently deleted.
Why it matters
This is the first major regulatory action where compliance requirements for AI agents are structurally incompatible with the product architecture itself — not just a matter of adding guardrails, but a fundamental conflict between persistent memory/identity and mandated disconnection. The regulation establishes a precedent that other jurisdictions may follow: when AI agents simulate personality and build emotional relationships, they become subject to anti-addiction and transparency requirements that can force platforms to choose between compliance and the feature's core value proposition. For the global AI agent ecosystem, this signals that agent governance is not just about security controls — it's about whether the product category itself is permissible in certain regulatory environments.
What to do
- If you use Doubao or Qwen agents, export your data immediately — screenshots and text exports before July 15. Doubao data is gone after October 15; Qwen data is gone immediately.
- For organizations building AI agent platforms with persistent memory, review your compliance posture against emerging anthropomorphic AI regulations — the China framework may influence EU, US, and other jurisdictions.
- Consider the architectural implications: if your agent platform builds emotional continuity through persistent identity and memory, you may need to design mandatory disconnection mechanisms from day one.
- Monitor regulatory developments — this is the first of what may be multiple jurisdictional approaches to AI companion/agent governance.