A social-media fad of hobbyist AI “shrimp‑raising” — users installing experimental agent tools like OpenClaw on personal machines — has shown early signs of cooling. DIY installers who paid hundreds of yuan for setup services are now paying others to uninstall their agents, and government and institutional warnings about security and privacy have followed. Yet while individual enthusiasm ebbs, China’s biggest internet firms have rushed in with polished, interoperable, cloud‑backed agent products: Tencent’s WorkBuddy, ByteDance’s ArkClaw, Alibaba’s CoPaw, Xiaomi’s miclaw and NetEase Youdao’s LobsterAI among them.
What looks like a chaotic scramble is, in fact, a coordinated, fast‑moving strategic contest. In a window of barely three weeks from late February to early March, more than a dozen companies completed product launches, cloud deployments and promotional pushes almost simultaneously. Superficially the moves are similar — one‑click deployments, prebuilt templates and user subsidies — but firms are pursuing distinct, position‑specific bets.
Cloud providers such as Tencent Cloud, Alibaba Cloud, Baidu Intelligent Cloud and Volcano Engine are selling the infrastructure that makes agents usable at scale. They productize deployment and management to convert complex local setups into standardized cloud services. The calculus is simple: models and agents consume compute and tokens; whoever supplies that compute captures the recurring revenue stream — the new “electricity” of the AI era.
Platform owners are fighting for the first user touch. Tencent’s WorkBuddy integrates into WeChat and QQ so agents can sit inside the chat windows where hundreds of millions of Chinese users already interact. ByteDance ties ArkClaw into Feishu and other workplace tools to ensure its agents become the execution layer for user requests. For these companies, being the place where users speak the first command matters more than arguing about who built the cleverest model.
Others are embedding agents into device ecosystems and developer flows. Xiaomi’s strategy is to fold agents into the system layer and AIoT family to make task automation a device‑level feature, while NetEase Youdao offers a simpler “small‑user” experience that replaces command‑line tinkering with installers and packages. Alibaba has open‑sourced CoPaw and pushed low‑cost developer plans not from altruism but to keep developer mindshare and the pre‑transaction commerce funnel within its ecosystem.
The strategic stakes map neatly to each company’s historical strengths. Tencent defends social initiation points and the moment users choose to act; Alibaba defends the commerce discovery and decision path that precedes payment; Baidu defends search’s role as the default starting place for information tasks; ByteDance defends its algorithmic distribution and the right to decide what users see. If agents let users delegate tasks by voice or text, those starting points may migrate away from today’s platforms.
That divergence explains why the market now resembles a contest for “ponds” rather than for individual shrimp. Early adopters who installed agents have learned a lot — they now have cloud accounts, payment methods and a basic literacy about tokens and model limits. Even when users uninstall experimental software, the habits and infrastructures remain. Tech giants are competing to own those durable assets: accounts, billing relationships, developer ecosystems, platform APIs and the privileged channel to accept user instructions.
The risks and frictions are real. Security incidents and privacy concerns have prompted cautionary guidance from institutions and even local bans in some organizations. The quick consumer cycle — pay to install, pay to uninstall — underscores how immature and sometimes hazardous early agent deployments can be. Regulators and enterprise buyers will look for safer, governance‑ready offerings, which advantages the cloud operators that can offer managed, audited deployments.
The immediate implication is that the “shrimp” craze may pass, but the contest for platform primacy in the agent era will harden. Whoever owns the cloud, the entry point and the distribution rules will capture the economics of task automation: token consumption, recommendation and ad value, and the privileged pathway to user intent. The next phase will be less about hobbyist creativity and more about durable commercial and regulatory plays.
For global observers, China’s sprint matters because it offers a preview of how vertically integrated platform ecosystems can shape agent development. The contest mixes open source, proprietary models, device integration, and cloud commodification — and the winners will set defaults that affect usability, privacy norms, and who captures the revenue generated when AI actually does people’s work. The “shrimp” may be ephemeral; the ponds they reveal are not.