After “Shrimp Farming”, How Does AI Agent Leap Forward?

A viral tool, a bigger shift

A suddenly viral open‑source project known as OpenClaw (龙虾) has done more than attract clicks — it has reframed expectations about what AI can do. Reportedly, users in China found that the project could tie together calendars, email and browser actions, turning conversational models into action‑oriented assistants. Is this the start of an “agent era”? The bigger story is not the meme itself but a shift in perception: AI is moving from “I can answer” to “I can act.”

China’s practical adoption and the “second‑class” AI

China’s market — hungry for single‑person productivity multipliers among individuals, enterprises and local governments — has pushed adoption in a pragmatic way: not through splashy launches but by asking, “Can it do useful work?” Reportedly, OpenClaw spread through installable services, tutorials, cloud hooks and enterprise pilots. But technologists caution that what looks like autonomy is often API orchestration: systems that recompose existing tools into smoother action chains. Products such as Manus and CoPaw have surfaced as variants — one making the execution pipeline visible, another stressing modular, multi‑channel deployment — illustrating that the immediate frontier is integration and transparency, not synthetic agency.

Who’s racing to the “third‑class” agent?

Some Chinese firms are explicit about pushing beyond orchestration. MiniMax — led by Yan Junjie (闫俊杰) — has publicly framed itself as moving from a model house to a platform company that aspires toward AI‑native agents. It has reportedly seeded proactive behaviours, multimodal inputs (text, voice, video, music) and internal agents across staff workflows. Yet limits remain: proactivity is still shallow, cross‑system real‑world action loops are incomplete, and a consumer‑grade legal and governance base is unfinished. In short, MiniMax shows the technical and organisational signs of a third‑class agent but not the full stack required for safe, autonomous operation. Kensho, by contrast, is cited as closer to true third‑class work in high‑value verticals where it can integrate multi‑source information and act within defined boundaries.

Risks and geopolitical context

It has been reported that recent coverage has flagged risks — from rapid cash burn and faulty actions to sensitive directory access, data leakage and attack vectors — underscoring that orchestration at scale raises safety, audit and rights questions. Add geopolitical pressure: export controls and chip restrictions from the U.S. have accelerated China’s push for domestic stacks and open‑source innovation, sharpening incentives to build locally controlled agent platforms. OpenClaw’s virality therefore marks more than a tool craze; it signals a competitive phase where success will hinge on practical integration, transparent governance and the ability to safely turn “say” into “do.”