NextFin News -- Over the past two-plus months, the open-source agent OpenClaw, nicknamed “Lobster” in China, has gone viral quickly. Alongside the tech frenzy, a wave of collective anxiety is spreading across the entire industry: big tech worries that traffic gateways will be reshaped, office workers worry their jobs will be replaced, and security experts—amid blaring alarms—worry that core system data will spiral out of control.
Why have the platforms that major tech giants poured huge money into over the past year generated only lukewarm responses, while an open-source “lobster” managed to break out of its niche and reach the mainstream? In a recent internal talk, 360 Group founder Zhou Hongyi offered a prediction for the industry: agents are shifting completely from “chat tools” into “digital employees” that can actually get work done—and that shift will directly upend the existing traditional software ecosystem.
In Zhou’s view, although China’s AI industry experienced a boom over the past year, large models were still more like chatbots with no hands. Most agents promoted by big companies remained stuck at the Q&A stage; in essence, they were still software tools built around preset, fixed workflows.
Lobster’s breakout, however, validated a “Scaling Law for skills.” Once an agent is granted extremely high permissions and freedom, it is no longer a program that only reacts when triggered. Instead, it becomes a new kind of productive force—able to explore on its own and combine thousands of tools to solve problems. It has helped the public truly understand what an agent is: a silicon-based worker with autonomous execution and self-correction capabilities.
As agents grow more capable, they are fundamentally reshaping the software ecosystem’s balance of interests. Zhou Hongyi argued that traditional software will move “down the stack,” becoming more plug-in–based and modular—turning into a “raw-materials library” that agents can call on at any time.
In the past, users could only passively accept hard?coded features and settings written by developers—like reheating ready-to-eat prepackaged meals. But Lobster Mode has broken that pattern. Zhou Hongyi predicted that programming will become a foundational capability for all agents in the future.
When it encounters a problem with no off-the-shelf tool available, Lobster can “code as it goes”: it writes a short piece of code to solve the problem and then discards it right away.
Despite its enormous productivity potential, rolling out OpenClaw faces an extremely high technical threshold. Setting up virtual machines, configuring development environments, and connecting to various APIs typically takes even seasoned engineers around six hours. At the same time, granting agents high-level system privileges introduces critical security risks such as prompt injection and leakage of core data.
To address these two major pain points, the 360 Group officially released the “360 Security Lobster” agent application client and the “360 Security Lobster Box” hardware terminal. This solution compresses the complicated configuration process to under 10 minutes, enabling true one-click, out-of-the-box use.
On the security front, 360 follows the “principle of minimum privilege” and simultaneously introduced a native security component, “360 Lobster Guard.” Through a virtualization sandbox (WSL) and physical isolation technologies, it defines security boundaries for government and enterprise customers as well as individual users—without interrupting agent tasks.
The following is the full text of a talk given by Zhou:
Lobster has been especially hot lately—everyone’s raising lobster. Today I’m not here to do a product launch; I want to talk about what’s really been happening behind this little crayfish that suddenly went viral.
For the past year or two, I’ve been talking about agents every day, but people felt it was all airy and intangible. Then Lobster appeared and broke out of the niche perfectly. What exactly did Lobster get right? The major internet companies have been going all out pushing their own platforms—so why didn’t they manage to ignite the broader public?
Because what we were building in the past was still tool-type agents. No matter which platform you used to assemble something, its functions and workflows were tightly locked down by us orchestrators. It isn’t a silicon-based employee; at its core, it’s still a piece of software running along a predetermined route.
But Lobster proved the “Scaling Law of skills.” When you give an agent three tools, it can’t do much; but when you give it 3,000 or 10,000 tools, it can mix and match them at will. It treats itself like a real person and learns through trial and error. You ask it to take a photo on a computer; it doesn’t have that capability out of the box, so it goes to GitHub, downloads an open-source camera-control program, gets it working, and sends you the result. That’s the ability to evolve autonomously. In the past, we were still building AI with an “app-making” mindset—and that was a conceptual mistake.
This leads to a very brutal reality: who is the biggest maker of “pre-made meals”? The traditional software industry.
Traditional software is essentially pre-made food. Developers “cook” the code at home, bring it to you for a bit of configuration—like a microwave meal: it is what it is. Take it or leave it; all you can do is wait for the next version. But with Lobster, selling software this way no longer has value. Future software will inevitably move toward being more foundational, plug-in based, and modular. Users won’t be buying a single massive software license anymore; what they’ll want is a bunch of Lobsters that can juggle all kinds of software skills.
In this process, one capability will be fully unleashed: programming.
People used to think programming was the exclusive domain of programmers. But going forward, programming will become a basic capability of all agents. When Lobster has tried 10,000 off-the-shelf tools and still can’t meet your request, what will it do? It will look up information online and write a piece of code on the spot. Write it when you need it; throw it away when you’re done. Software costs will be squeezed to extremely low levels. In the future, programmers may no longer need to type out business code line by line themselves—they’ll only need to focus on properly training and steering the Lobsters under their command.
Of course, as Lobster took off, some people began to worry about security.
Where there’s power, there are risks. If you hand your safe key to an assistant, you’re bound to worry the assistant might misuse it. Agents can hallucinate, be manipulated by people with bad intentions, and suffer injection attacks. Many people think Lobster is a virus, and the moment they see it touching the C drive they want to block everything. But what I want to say is: the greatest insecurity is refusing to develop and refusing to move forward.
When we were building security solutions, we had intense internal debates: should security come first, or should the business come first? In the end I made the call and adopted the “minimum principle.” As long as the most dangerous kinds of catastrophic damage don’t occur, I can tolerate it. If the moment an agent accesses the camera you immediately pop up an alarm and forcibly terminate it, then Lobster won’t be able to do anything at all. What we need to do is lock down the most critical things—online-banking passwords, API keys—give it a safe virtual environment, and treat it as a capable assistant, not guard against it like a thief.
Lobster is only the beginning. If you treat it as a smart intern—embracing its trial and error and letting it grow—you’ll find that a whole new era of business has already begun.
快報
根據(jù)《網(wǎng)絡(luò)安全法》實名制要求，請綁定手機號后發(fā)表評論