A conversational craze has taken hold in China’s tech scene: not shellfish, but software “lobsters” — autonomous AI agents built on the open‑source OpenClaw project that can operate desktop applications, send messages and run workflows without constant human direction. On March 13 NetEase hosted a high‑profile salon in Beijing called “Bring You to Play OpenClaw Lobster,” gathering CEOs, researchers, product leads and power users to turn the trend from viral anecdote into practical playbook.
Speakers outlined why interest has mushroomed. OpenClaw lowers the engineering barrier to giving large models hands and feet: agents can click, type, open apps, orchestrate APIs and persist state. That capability converts a generative model from an assistant that answers prompts into an independent executor that can manage email, draft posts, run data extraction and even supervise repetitive business tasks.
The salon mixed evangelism with hard practicalities. Industry figures described deployment paths — local machines, cloud images and integrations for workplace chat platforms such as Feishu, WeChat and DingTalk — and compared trade‑offs in cost, performance and security. NetEase Youdao’s LobsterAI, released in February 2026 as a desktop 7×24 personal assistant, was showcased alongside commercial and hobbyist installations to illustrate how agents are moving from experiments to day‑to‑day utility.
Adopters recounted striking productivity gains and surprising uses. Fu Sheng, chairman of Cheetah Mobile and OrionStar, said his “lobster army” runs around the clock to manage work, coach his rehabilitation exercises and even update his public account, spending more than RMB 30,000 a month on model tokens and working into the small hours. Other presenters demonstrated automated product‑design workflows and entrepreneurial setups in which one person leans on agents to run substantial operational loads.
The economics are double‑edged. Tokens and compute can add up quickly; some users report daily or monthly bills that make indie adoption expensive, while hardware sellers and cloud providers have seen a surge in demand. Municipal governments are responding with incentives: a number of localities have unveiled subsidies and “dragon‑shrimp” (longxia) policies to seed startups and single‑person companies built around agent tooling.
But the same capabilities that make agents powerful create new vulnerabilities. With privileges that reach into operating systems and business accounts, agents can exfiltrate data, escalate privileges or be subverted by prompt‑injection attacks. Salon panels emphasized sandboxing, principle‑of‑least‑privilege, observability and cost controls as essential countermeasures. The industry’s shorthand for catastrophic failure — “炸虾”, or frying the lobster — summed up concerns about runaway costs and security breaches.
Beyond consumer excitement, speakers argued agents signal a structural shift: from model‑centred competition to systems that bundle models, connectors and orchestration. OpenClaw’s open architecture accelerates experimentation and lowers entry costs for “super‑individual” entrepreneurs, while companies are racing to productize agent frameworks. That creates opportunities for tooling vendors, cloud providers and enterprise IT teams, but also concentrates risk if insecure agent deployments proliferate at scale.
The salon illustrated a tension common to frontier AI technologies: rapid utility and commercialization, paired with immature governance. For international observers the Chinese episode is instructive because it compresses several diffusion dynamics — open‑source tooling, consumer memes, municipal policy and enterprise product moves — into a single boom. How firms and regulators respond, and whether standards and sandboxing keep pace with adoption, will determine if these agents are a durable productivity advance or an expensive, insecure fad.