A weekend project by an Austrian developer became the fastest growing repository in GitHub history. It triggered $1 trillion wipeout in software stocks, caused Apple Macinis to sell out globally, moved Cloudflare stock 14% and prompted a Chinese district government to build industrial policy around it. This video is about the next big thing in tech, agent operating system, and we're going to use the most recent precedent, the company called OpenClaw.

Since the inception of the company, OpenClaw went through a triple rebrand. The original name was Claudebot, then renamed to Moldbot, and then to OpenClaw. But none of that naming nonsense slowed down the adoption. 25,000 GitHub stars in a single day. That's the highest daily count ever recorded on GitHub. The fastest repo rise in GitHub history. 250,000 stars by March 4th, 2026. and surpassing React which took over a decade. Open Claw is not an AI model and it is not an app. It is an orchestrating layer that sits between your messaging apps and any large language model to perform a bunch of tasks. Be it manage email, check into flights, organize files, hire people, fire people, fill out documentation on your behalf, etc. On a very high level, open clause architecture works like this. It uses messengers like WhatsApp or Telegram or Discord or Slack as its interface. You run your tasks and commands directly from the messaging app. A message is then routed to appropriate agents. Then it gets converted into a prompt and then passes that context to an LLM be it Claude, GPT, Gemini, Llama, whatever you pick and executes the command. Now why is this a big deal? Every major computing era had a defining milestone. It had a foundational layer that defined it. And foundational layers often look unimpressive at first. Databases, for example, back when they appeared were a massive overkill because you could write custom code to handle files. Browsers back in the day when they first came out were much slower and uglier than native or on-prem apps. Nevertheless, each became a default piece, a default infrastructure because it solved problems that other applications and other businesses were built upon. For example, Windows and Mac OS were the milestones of the '9s and 2000s. Chrome was and still is the absolute leader of the browsing market. AWS, Azure, and Google Cloud define the cloud. LLMs, computer vision, genai are equivalent to OpenAI, Anthropic, or DeepSeek. And now we're witnessing the birth of the next foundational layer, agent OS. The agent OS is essentially a new browser war, but the stakes are an order of magnitude higher. The appearance of Open Claw and how it exploded is a sign of a fundamental UX shift in artificial intelligence. Once again, it's a UX shift because architecturally, Open Claw is not a mind-blowing discovery. The whole thing was vibe coded in a couple days. This is a UX shift because the realization that agents can bypass multiple software user interfaces and interact with APIs directly challenges the assumption of the past 30 years. For 30 years, every piece of software was built around the assumption that a human would use it. A human would open it. A human would look at the screen, click the buttons. Careers, companies, consultancies were built around the assumption that there is a piece of software that you learn and you sell that skill for a living. Everything you see on the screen right now is months of user research, design, and development. What open claw is doing is that it bypasses the user interface layer of multiple apps. Instead of going and performing actions by hand, you give the agent a goal. For example, let's use a business case scenario. Let's say we go update all open deals in Salesforce where the last contact was over 30 days ago and flag them as stale. Did you notice that I did not do this in Salesforce? I didn't log in. I didn't use the app. Salesforce spent months or years building this interface for me to go in and do these things quickly, but I am at a point where I don't need this interface at all. This is why it's a UX shift. Just like the buzzword of the last year was AI agent, Open Claw is what comes after the agent. Open Claw is the distribution center of agents. You're not bound to any particular LLM. You're not bound to an app. As a casual user, you don't even care how your requests get parsed, be it through Claude or GBT or Gemini. The point is, you're not using the app anymore. The app that was built and refined over the years, you're not using it to get things done.

So, after all of this, does this mean that days are numbered for the software market? No. No, it doesn't. There are numerous conversations happening about the death of the software app market. On February 3rd, around $285 billion were erased from the global software stocks in a single day. Some of it came back, but most of it didn't. And yes, we always think that stocks always come back, but the market confirms that something has fundamentally changed. Claims around the death of SAS are primarily based on the assumption that AI agents become the primary way that we will interact with software. But the fact that you can route your tasks to agents does not mean that you will. Within the B2B SAS world, which ironically I spent a decade in, the prevalent pricing model has always been per seat. And yes, it's painfully obvious. The higher the number of people they use a piece of software, the more the customer pays, the more your business makes. Can an agent break the paper seat model? Yes, very much so. An agent can simply compress the number of seats that a customer would need. For example, let's say I'm a midsize office furniture business and I have a team of 10 sales reps using HubSpot. They go on HubSpot, they build a pipeline of prospects, they fill out their contacts notes, and this takes 30% of their day, and only 70% are spent on selling the actual product. And now, Open Claw comes in and says, "I'm going to give you one sales agent and use HubSpot's API." So, instead of paying for 10 seats, you're paying for one agent, connecting to HubSpot, and filling things out 24/7. Do you see the difference between a co-pilot model and this? A co-pilot would be sitting inside HubSpot. It would exist in a form of a chatbot, for example. You would plug in your prompt into HubSpot Copilot and say what you want done. This may or may not be a paid feature. You may or may not need additional credits to run this chatbot. Point is, it would be a HubSpot native co-pilot. A HubSpot would build it and HubSpot would own it. OpenClaw is doing exactly the opposite. It sits above your apps. It doesn't need a chatbot or a co-pilot and therefore it doesn't need the seats. Another example, think about how a marketer runs a Q1 marketing campaign today. They open Google Analytics to pull traffic data. Then they switch to HubSpot for lead scores and then they jump to Hootsweet for social scheduling. At the end of this process, perhaps build some email sequences in Mailchimp. Each tool has its own seats, has its own login and its own dashboard. And now imagine telling a single agent, plan my Q1 campaign, analyze last quarter's performance across all channels, generate content variations, schedule posts, set up AB tests, and optimize in real time as you get the results. The agent talks to each tools API, pulls the data, does the work, and reports back. On top of it, there's a whole layer of network effects that an agent creates. And it does so because it accumulates history. It would collect cross app insights and knowledge graphs that no single copilot can match. Until today, the vast majority of the software businesses were built across two layers. The backend layer, APIs, databases, business logic, and the front end or in other words, the user interface. What open claw can do. And let me emphasize that the operative word here is can do is that it can eliminate the need to have the top layer. And that is why you see the headlines about the death of SAS. So coming back to the market reaction in February when the stocks fell, SAS is not dying but it is getting hollowed out from the top where the interface layer is no longer the strongest selling point and we will be seeing a massive shift towards investments in the data layer.

The second reason why the software market is not dying is because there is one fundamental problem with the agents. Agents are eyewateringly expensive to run. And this is where we hit a wall. The fact that an agent can order an airplane ticket for you does not mean that you will use an agent to do that. It's one thing to say how fascinating that would be and another thing to pay for that fascination. I don't know if many of you have noticed, but open claw ain't cheap. software market is nowhere near being dead for one simple reason. Agents, as marvelous as they are, come at a very a very high cost. In addition to a bunch of failure modes that traditional apps with user interface do not have. Anyone who has built at least one agent for personal use will tell you how expensive it can get. Speaking from experience, I tried building a research agent for our channel in N88 to do the research faster as well as to get good at building agents. And the amount of money that I spent on API credits to connect various nodes to my agent was simply unjustifiable. And mind you, I was building it for personal use. Imagine scaling those costs to an enterprise level. the costs become astronomical and any employee with the power to make those purchasing decisions will think twice before signing up for a team plan to run those agents. Running autonomous AI agents is not free. It is structurally more expensive than a standard LLM query and the costs multiply almost exponentially as your agent becomes more advanced. Do the math. With agents and an agent orchestrator, you're multiplying inference. Every task expands into multiple LLM calls, a planning call, an executor call, error checks, retry loops. Instead of paying for one inference, you've got a compound cost graph that keeps scaling. You're also inflating your context window because agents maintain memory and they do need long context pricing tiers. And lastly, the execution costs because every tool call includes LLM reasoning cost, the execution costs of external systems, LLM synthesis cost, and API tokens across multiple applications that you call. And to top it off, multi- aent systems need agent to agent chatter. And what that means is a spiraling token cost. So once the initial fascination fades, we're going to hit a wall with these costs. My point is basically this. The fact that you can get an agent to order and plan a whole trip for you with all the bookings and restaurants and hotels and museums doesn't mean that you're going to use that agent to book all of your future trips for the rest of your life. Because if I tell you that booking a trip in that manner is going to cost you $40 and that gets multiplied by the number of tasks that you're outsourcing to agents, clicking on various buttons seven times to get that ticket ordered doesn't seem too shabby. So no software market is not and will not be dead.

What's really important to observe in the coming months is who becomes the dominant player in the agent OS market because whoever dominates the agent operating system market gets at least three massive control points over the rest of the tech industry. This is a power play that is unfolding. The first point of control is going to be identity and authentication. Now, why is this a big deal? Prior to the rise of agents, identity in the context of software meant that one person would be logging in with a username and a password. And yes, the process of logging in can take many different shapes. But every access management system was built around the assumption that it would be a human logging in. And access management is a 20 billion industry that is dominated by Microsoft, Octa, and Cyber Arc. When agents enter the picture, the question that gets raised is who becomes the default identity provider for the agents. Microsoft already dominates the market for human authentication through Active Directory. So if Microsoft Copilot for example becomes the default agent identity inside Microsoft 365 every enterprise that uses M365 which is essentially all of them now needs Microsoft's permission to run any agent. That is the level of leverage that Microsoft can get because of this market situation. The second control point is data routing. Agents spend all day moving information between various systems. They're pulling customer data from Salesforce, cross referencing it with email history in Gmail, updating a project board in Notion, and then summarizing and spitting out a message on Slack. Every one of those information flows passes through the agent OS, which decides what data goes where and in what order. The agent operating system sees the behavioral map of every single user. They see that a company X is hitting their legal database 40 times more than the last month, so something must be brewing. Company Y is actively adding severance information to their HR system. Could it be that they're planning layoffs? And Company Zed is pulling competitive pricing data every single hour. Are they repricing? This data is more valuable than any market research you can imagine. Every major data broker, hedge fund, consultancy would pay extraordinary sums of money for this behavioral intelligence. Will Agent OS monetize this data directly or sell access to it indirectly? Who knows? But let me remind you that Google built a $300 billion a year advertising business on knowing what people search for. So there's that. And lastly, probably the most commercialized consequence of this entire situation is the default model selection. When your agent performs any kind of task, it calls an LLM. And it is the agent OS that decides which model gets called first and gets set as a default. For the vast majority of casual users, they will never change that default setting. Global spending on AI inference is projected to reach $200 billion within the next two years. And if the dominant agent operating system controls routing for even 30% of B2B enterprise traffic and sets the default model to one where it has a revenue sharing arrangement, that's approximately a $60 billion revenue coming from a single setting. For comparison, the entire Google Chrome, one of the most valuable defaults in computing history, makes $20 billion a year from search traffic defaults. The agent OS has a financial incentive to route traffic to a specific model in case it has a revenue sharing agreement with the company that produces the LLM. Now, does it mean they'll do it? No, it doesn't. But they can. If clot is objectively better, let's say 12% better at legal reasoning than OpenAI, but OpenAI is paying someone like OpenClaw 15% revenue share. It could route legal tasks to GPT anyway. Now, who benefits from this? OpenAI is already the beneficiary. They have the most popular consumer model on the market. They now employ the creator of OpenClaw and they have the distribution through Chad GBT. Enthropic on the other hand is pushing the model context protocol as a standard for how agents connect to tools and they're doing so to prevent any agent operating system from having routing control over which model gets called. Meanwhile, Google is embedding Gemini so deeply into Android and workspace that the agent layer on Android devices defaults to Gemini with or without third party routing. Apple is doing the same with Apple intelligence on iOS. What we're seeing is that every major player in tech is racing to own the agent OS default. Now, everything I just described is the war playing out in the US.

But there's a parallel story happening on the other side of the world that is frankly just as fascinating. Let's talk about China. A local Chinese government built a formal industrial policy around a piece of open-source software created by an Austrian developer by an American company running on US servers. The district that built this policy is called Long Gang, and it just happens to be the home of Huawei. The fact that Huawei's home district published a policy around an American open-source agent framework while Beijing simultaneously warns state enterprises not to install it on their office computers is perhaps the most mind-boggling paradox in this whole situation. Chinese companies are running OpenClaw as the agent orchestration layer using Chinese domestic LLMs on local Chinese hardware, communicating through Chinese apps and subject to Chinese data law. The Chinese central government and local governments are pulling in opposite directions on OpenClaw and neither side is blanking. On one side, Beijing's national regulators have warned state-owned companies, major banks, and government agencies against installing Open Claw on office devices. On the other side, local governments in Shenzhen and Hafe are actively subsidizing the adoption with millions of dollars in public funding, running training sessions on how to use OpenClaw, and explicitly including OpenClaw in their national AI plus plan. And the irony of this is that this dual track strategy that China is exhibiting has been done before. Linux in the 2000s, Android in 2010s, and now the agent OS. They let the local market adopt the foreign open-source framework. They built domestic alternatives and forks in parallel. They apply the foreign framework everywhere except the state apparatus itself. They fully adapt to the local market and continue iterating on top of what they learned. They utilize the best of the western tech to make it even better. We're going to be talking a lot more about China, but that's going to be in the next video.

Will OpenClaw be the company that defines the age of agent OS? Personally, I don't think so. The app has many flaws. Its security is raising a lot of questions. The thing was vibecoded to begin with, so calling it a generation definining flagship product would be an overstatement to say the least. But the thing is, it doesn't need to be the Windows of the decade. What just happened isn't about Open Claw. It's about the signal that it sends. What it's telling us is that the interface that every software company spent the last 30 years perfecting is no longer the moat. The new mode is sitting above the interface in the identity and in the data. As always, we hope this was helpful. We'll see you next time. Bye.