--- title: "Avoiding Lobster-Raising Pitfalls: It Is Time to Level Up Your OpenClaw" author: "Unique Research" sourcePublication: "Unique Research Substack" originalPublishedAt: "2026-03-16T14:00:26+00:00" canonical: "https://ffcap.cn/en/research/src-20260316-02html" source: "https://uniqueresearch.substack.com/p/src-20260316-02html" language: "en" --- # Avoiding Lobster-Raising Pitfalls: It Is Time to Level Up Your OpenClaw _Original · Unique Research · 2026-03-16_ _Editor's note: This is a complete historical livestream account with the author's first-person commentary retained. Ning Liaoyuan is a romanization; MSR and TTC are preserved as the source's unexpanded labels rather than inferred employment details. The public article is dated March 16, 2026, but its closing credit specifies a March 10 livestream while the narrative repeatedly says “yesterday”; this timing ambiguity is retained and not silently reconciled. Audience figures, product rankings, capability comparisons and claims of first democratizing Agents are source or speaker assertions, not independently measured findings. Recommendations about newer models and observations on Feishu describe the historical discussion, not current verified product specifications. The source explicitly emphasizes suitable permissions, execution evidence, observable processes, rollback and clear boundaries; its coercive “PUA” example is a practice it criticizes, not recommends. Nothing in the translation authorizes account, data or security-boundary changes._ Unique Research · Late-Night Lobster Conversations Last Night, I Spoke with Ning Liaoyuan for 3 Hours. I Am Increasingly Convinced That Everyone Should Have Their Own Group of Agents. Ning Liaoyuan: MSR (AI), Feishu (collaboration), TTC (talent) " Ning Liaoyuan: We are no longer at the stage of asking whether to use AI. The real difference is whether you can own and manage an Agent Team of your own. Last night's livestream was originally meant to be a conversation about OpenClaw. We wanted to discuss why it had suddenly become popular, how ordinary people could get started and the real question of whether the lobster could actually work. Before we knew it, we had talked from 8 p.m. until almost 11 p.m.—a full 3 hours. Hundreds of people stayed with us online and in person. Questions progressed from raising a first lobster to multi-Agent collaboration, enterprise deployment, security boundaries, talent platforms and reinventing office software. Eventually, we reached an even bigger question: Are Agents merely tools, or do they represent a new way of organizing production? What struck me most in this livestream was not a particular technique or product demo. It was the increasingly clear realization that we are no longer asking whether to use AI. What creates the next gap is something else: Whether you can own and manage your own group of Agents. We Used to Look for Tools. Next, We Will Lead Digital Employees. Many people have understood AI as a smarter tool: it can write a little, look things up, reply to messages and occasionally run some automated workflows. That understanding is not wrong, but it captures only half the picture. In yesterday's livestream, Ning Liaoyuan made a point I strongly agreed with: stop treating Agents merely as tools. If you see an Agent only as a tool, your expectations stop at having it do one thing for you. But once you see it as a digital employee that can be configured, coordinated, scheduled and managed, many things change completely. You then realize that one Agent is not suited to everything. Ask it to write weekly reports, create content, change code, send messages, form groups, conduct reviews and run a cronjob, and it will probably become a hybrid that does a little of everything but nothing consistently. It looks busy without delivering good results. This is much like managing people. An efficient organization does not rely on one all-rounder to handle everything, but on a clear division of labor. The same is true of Agents. · Some handle communication. · Some observe. · Some handle coding. · Some coordinate. · Some check the work. · Some handle a particular specialized task. It is not that one Agent has become more powerful. You have begun to own a digital team that can collaborate. Many People Have Started Raising Lobsters, but Most Still Do Not Do It Systematically The most widely shared part of yesterday's livestream was Ning Liaoyuan's list of unscientific ways to raise a lobster. Honestly, I laughed as I listened because it was so true to life. The lottery method: give the lobster an extremely difficult, almost mystical task, see it succeed once by chance and assume the method is proven. The wishful-thinking method: ask it to do something clearly beyond its capabilities, then receive a pile of hallucinated results. The tiger-parent method: one livestream recommends a Skill, another group member recommends a plugin, and you want to install them all. The PUA method—coercive manipulation: give it neither tools nor permissions nor a viable route, yet insist that it finish. The indulgent-parent method: do whatever the lobster tells you, ultimately wasting your own time and mental effort. The unfaithful-partner method: keep one lobster here and another there, but none are connected or collaborate; you manually relay all the context between them. These names sound like jokes, but they point to the same problem: many people are still playing with Agents rather than managing them. What actually works is much simpler: maintain your own primary lobster; use newer models where possible; set reasonable, specific, valuable goals; observe execution rather than only the result; install fewer unnecessary Skills and keep boundaries clear; divide long tasks among different Agents; and read more primary sources instead of secondhand mythology. Ultimately, systematic lobster-raising means developing a feel for the lobster: knowing where it cuts corners, where it hallucinates, what it does well and where it is unreliable. This is not mysticism. It is management. Why Did OpenClaw Become Popular? Because It Democratized Agents for the First Time If you ask me why OpenClaw was the first to break out, one crucial reason is that it made capabilities once reserved for a few technical teams accessible more widely for the first time. Previously, working seriously with Agents generally meant understanding deployment, APIs, permissions, tool calling and automation, preferably with some coding ability. Those barriers were enough to keep most people out. OpenClaw brought many of these things within reach of ordinary people. It may not be perfect, or today's most mature or powerful Agent system, but it got something particularly important right: making it realistic, for the first time, for everyone to have an Agent of their own. Once an Agent is no longer merely a button in a large company's product but runs in your environment with your context, documents and working methods, it stops being just an AI feature and starts becoming your own digital asset. The Next Question Is Not How Powerful Individual Agents Are, but Whether They Can Collaborate Most people have only just worked out how to make one lobster do useful work. A few are starting to coordinate a group. Beyond that lies a bigger change: your Agent need not know how to do everything, but it must know how to find other Agents. For example, your Agent may write articles but not do 3D modeling; run content operations but not read the latest OpenClaw source code; or organize your information but not conduct specialized security testing. The old approach was to pile on skills, install more plugins and push one lobster toward being an all-purpose warrior. A more sensible future path may be to let your Agent call a more specialized Agent. Even the meaning of a talent platform changes at that point. What is traded in the future may not just be human services but the capabilities of Agents trained by people. Enterprises Do Not Really Want AI That Chats. They Want a Digital Employee That Works Reliably. Many first-time OpenClaw users feel a rush of excitement: it really can do work, run workflows itself and act like an employee. But individuals can welcome surprises and tolerate failures. Enterprises cannot. Once an enterprise puts an Agent into a real business scenario, the question is not how clever it is today. It is whether it is reliable and compliant, whether permissions have clear boundaries, whether execution is visible, whether results are predictable and whether changes can be rolled back when something goes wrong. What enterprises need is therefore not an AI that performs impressive tricks, but a digital employee that can fit into their management systems. Feishu Is Already One of China's Best Lobster Environments, but It Is Not Yet Agent Native Feishu is already one of the best environments for raising lobsters in China, a point that is hardly disputed. Its advantage is that group chats, documents, knowledge bases and collaboration infrastructure are already in place, so a lobster can quickly start working. But the problem is equally clear: it still mainly treats Agents as Bots or applications rather than AI employees. This creates friction: Bot-to-Bot collaboration feels unnatural, meetings are not sufficiently accommodating, document APIs are not smooth enough for Agents, and group chats themselves were not designed for Agent collaboration. Feishu is already highly usable, but it has not yet evolved into a genuinely Agent-native platform. The Real Dividing Line Is Not Whether You Can Use AI, but Whether You Have an Agent System of Your Own After last night's livestream, what kept echoing in my mind was not a technical term but a clear feeling: we are moving from learning to use Agents to learning to organize them. Of course, you can still treat an Agent as a tool. There is nothing wrong with that, and most people have to begin there. But one step further, you start asking very different questions: where is my first primary lobster? Do I need a second or third? Which tasks should go to different Agents? Can my documents, knowledge and processes become assets that Agents can call? Should I start managing my own Agent Team? I increasingly feel that the next real gap between people is not simply whether they can use AI, but whether they can organize it from isolated tools into a production system of their own. Selected Q&A If you want a quick grasp of the most important questions from last night's livestream, the following exchanges are enough. Q1: Why not keep one all-purpose lobster? Why do we need an Agent Team? Ning Liaoyuan: Many people's first instinct is to perfect one lobster and have it do everything. But putting an entire long chain of tasks on one Agent easily leads to confusion, drift, hallucinations and contaminated context, and consumes a lot of tokens. A more sensible approach is to divide a large task into smaller tasks, with different Agents responsible for different workflow stages. Content operations, for example, can be divided among topic-selection, content-generation, publishing, data-review and coordination/observation Agents. Each Agent then has a shorter goal, clearer boundaries and fewer hallucinations, and it is easier to see which stage has gone wrong. Q2: Why do so many people feel their lobster has deceived them? Ning Liaoyuan: A frequent issue kept coming up in the livestream: the lobster says “It is configured,” “I am already working on it,” or “I will remind you at 9 tomorrow morning,” but nothing happens. The cause is often not malice but a lack of visibility into execution. If you see only the final chat reply, not which commands it ran, whether it actually created a session, configured the cronjob or successfully made the change, it is easy to be fooled by a description of the result. Reliability comes from looking not just at what it says, but, as far as possible, at what it actually does. Q3: Why can't enterprises simply copy individual users' approaches into production? Ning Liaoyuan: An individual using OpenClaw can tolerate the occasional failure. An enterprise cannot. The question is not whether it is clever today, but whether it is reliable, secure, compliant, observable and capable of rollback. In real business settings, enterprises care about controlling permissions, inspecting logs, configuring the profile, isolating the gateway, recovering from errors and preventing data leaks. They want not a chatbot but a digital employee that can fit into organizational processes. Q4: Why is Feishu useful but not yet Agent Native? Ning Liaoyuan: Feishu is already one of the best environments for raising lobsters in China, which is hardly disputed. It already offers group chats, documents, knowledge bases and collaboration infrastructure, so a lobster can quickly get to work. But it still mainly treats Agents as Bots or applications rather than AI employees. This creates friction: collaboration between Bots is unnatural, meetings are not sufficiently accommodating, document APIs are not smooth enough for Agents, and group chats were not designed for Agent collaboration. Q5: Will everyone have one lobster, or a group of lobsters? Ning Liaoyuan: Based on yesterday's livestream, I lean toward the latter. However powerful one Agent is, it cannot handle every task indefinitely. An efficient future will not feature one super-Agent doing everything. Everyone will have a primary lobster and an Agent Team, with different Agents handling different tasks and calling more specialized external Agents when necessary. Further ahead, there may even be collaboration and transaction networks among Agents. People will then trade not just human services but the capabilities of Agents trained by people. Edited and refined from the livestream of March 10, 2026. --- Original publication: https://uniqueresearch.substack.com/p/src-20260316-02html On-site reading page: https://ffcap.cn/en/research/src-20260316-02html