Only 45 Days — Has "Longxia" (龙虾) Already Imploded?

Rapid rise, rapid unravel

OpenClaw — popularly nicknamed "Longxia" (龙虾) — exploded onto China’s AI scene at the start of the year as an open‑source agent that could take over desktop tasks, read files, call tools and run scripts. It reportedly became the most‑starred project on GitHub and prompted cities such as Shenzhen, Wuxi and Changshu to spruik policies supporting OpenClaw applications and new one‑person company (OPC) models. But the honeymoon was brief. In under 45 days the conversation shifted from opportunity to alarm: users, regulators and markets all pulled back almost simultaneously.

Users found it costly and risky

The immediate complaints were blunt: token burn and safety. It has been reported that some users saw hundreds of thousands to a million tokens drained after a few short interactions, and others said a simple web crawl cost nearly $20 in API bills. Even more worrying were safety incidents: users claimed that an abnormal Skill install once led to local disk contents being wiped — a claim that should be treated as unverified but emblematic of broader fears. In response, "500 yuan home install" help services were followed by lower‑cost "299 yuan remote uninstall" offers as DIY experiments turned into remediation chores.

Market jitters and regulatory warnings

Markets reacted. Shares of companies tied to the "Longxia concept" moved sharply — Zhipu (智谱) jumped when it launched OpenClaw on its platform, then fell the next day; MiniMax also declined. Regulators weighed in: the Ministry of Industry and Information Technology (工信部) flagged high security risks in default or mis‑configured OpenClaw instances on March 8 and cautioned again on March 11 that patches did not eliminate all threats. Feishu (飞书) CEO Xie Xin (谢欣) warned on social media that running agents on a personal PC is exploration, while enterprise deployment is responsibility.

Why the problem is structural

Experts point to the architecture. Qveris.ai COO Qu Dongqi (曲东奇) told reporters that OpenClaw’s full‑chain agent workflow—breaking goals into sub‑tasks, multi‑step reasoning, tool calls, state checks and carrying system prompts, identities and memory into each context—drives token usage far above single‑turn chat models. Ciling Technology (缔零科技) CEO Tan Yilang (谭亦朗) warned that simple file or memory tampering can create persistent cognitive backdoors in an agent’s behavior. Zhipu’s AutoGLM head Liu Xiao (刘潇) notes the framework’s plugin flexibility raises the deployment bar for non‑technical users and that teams are adding safeguards, but open‑source flexibility also means risk scales quickly. So has the agent era fizzled out? Not yet — but Longxia’s first month exposed painful trade‑offs between autonomy, cost and safety that will shape China’s next phase of agent adoption, especially as domestic AI policy and international scrutiny push more workload to local, offline systems.