But the side that actually affects us more directly is the enterprise adoption of AI, and that means your AI co-worker may have just been hired, and likely many more will be hired in 2026. And while I will congratulate you on your attempts at eliminating the presence of an AI companion in your personal devices, you will not have a choice when a company hires the AI that becomes your future replacement. Some of you repeatedly state in the comments that AI is not ready, that AI is dumb and how it will always fail. But if you believe this, you’re not understanding what Microsoft is up to.

Microsoft Windows now calls itself an agentic OS, meaning it is made to run AI agents. Microsoft is betting billions on this. And many of you don’t even know what an AI agent is as compared to the LLM that you have heard of, like chat GPT, Google Gemini, Anthropic, or Grok. There’s a lot to unravel here, and this is a warning video to all. You have to be smart. You have to understand the technology so you’re not blindsided. Stay right there. First, let’s clear up the confusion. Most people know large language models, LLMs like chat GPT or Grok.

These are basically super smart text predictors. You give them a prompt, they generate a response based on patterns from massive training data. They’re great for chatting, writing emails, or brainstorming ideas. But they stop there. It’s a chatbot. They don’t do anything on their own. AI agents are the next level. They’re built on top of LLMs, but add autonomy, planning, and action. An agent doesn’t just talk, it acts. It can break down complex tasks into steps. Use tools like browsing the web, accessing files, running code, or integrating with apps. Loop and iterate. Observe results, adjust plans, and keep going until the goal is achieved.

Remember, contacts over time and learn from feedback. Think of an LLM as a knowledgeable advisor who gives you advice. An AI agent is that advisor plus an executive assistant who actually books the flight, send the emails, and follows up, all without you micromanaging. Microsoft calls this agentic AI, and they’re getting big on it in the workplace. Agent autonomy. Many of you have experience working with AI chatbots, and I’m sure that as a personal user, you find it hard to rely on this unless you’re a student using it to do your homework. It just appears like fluff.

An AI will appear like that because a chatbot is not independent. It doesn’t do anything other than respond to your questions, and you have to prompt it each time. But AI agents do not need instructions. They can work autonomously. An example of this is Windows Recall, which is an autonomous AI module that can work full-time, taking screenshots every few seconds, and then analyzing what you’re doing and storing it in your personal history log. Very invasive, but very independent. And this is also the important reason that I gripe about AI because these agents are not under your control, particularly when you didn’t install them.

And that’s why they’re the worst privacy offenders in the modern OS era. LLM plus AI agents. What changed recently, and what’s advancing this fast evolution of AI, is that AI agents can now be linked to an LLM. So you can just assemble an LLM with instructions for command structure, and then the LLM can start issuing tasks to a collection of AI agents. So this moves away from just an AI chatbot. Now, the AI can do things. The biggest mistake is to assume that you need the smartest of all AI to figure all of this out.

In fact, without any further improvement in LLMs or even using free open source models, it can already perform tasks suitable as your co-worker. Modular agents. Unlike in the past, when so much programming was involved in attaching tasks to AI, this is now simplified. Anthropic, the creator of Claude LLM, was one of the movers in 2024 with the publication of the Standard Model Context Protocol, or MCP. MCP is nothing particularly complex. It’s basically just a wrapper to normal API calls that an enterprise will already have. But as an MCP server, it can accept commands from an LLM.

The LLM can continuously monitor the actions of the MCP server and then redirect actions to other MCP servers. MCP is probably the biggest buzzword in AI agent modules, and this is because Microsoft itself directly supports the installation of MCP AI agents in Windows 11. This means that a Windows computer can orchestrate the actions of multiple agents together with the understanding of an LLM to process actual tasks. I’ll give you an example in my livestream of insurance claims processing. An insurance claim has to be processed at multiple levels before it gets assigned to a claims adjuster.

Someone has to check policy validity, see if the claim is covered, and that could be an AI agent. Another AI agent could be determining fault and coverage limits. Another AI agent could be looking at the property damage and be arranging estimates, and so on. And maybe a final agent is in charge of final payments and then that is then sent to a check processing agent. While an AI could tell you what to do on the claim, an LLM is not task driven, it just talks. Now we have programs that do the work, and to be honest, much of it is just repurposing existing APIs in some format like MCP.

There are other formats, some AI companies who specialize in an industry like shift technology which focuses on insurance may have come up with their own internal format for communicating with an AI agent. Even Anthropic itself came up with an alternate agent wrapper called a skill with folders and a Python script. But for new companies, it is a lot easier. Microsoft Azure has a list of MCP servers you can rent. Amazon AWS has MCP servers, Google does, and so on. So basically because MCP is a common format, it can be used by any LLM that has some MCP client capability, which means all the major ones.

Microsoft is betting on the agentic OS. Microsoft is betting big on the agentic OS. As I stated many times before, Microsoft makes most of its money from Microsoft Azure, its cloud services, and that’s where the Windows co-pilot infrastructure is housed. What can happen here in a typical enterprise is that some can create agents that do data intake, for example, if physical documents are involved, or photos. There are MCP servers that handle computer vision and OCR that can describe what’s in the photos. MCP servers can also validate people’s photos using facial recognition.

Again, one of the capabilities offered by AWS, for example. Corporate knowledge bases can be kept locked up inside a corporate network, but channeled through MCP servers to provide the rule base for an agent to follow. The point is that when you have an OS that can directly handle MCPs, and you can run a Windows computer as a virtual machine in the cloud, then suddenly you have employees you can spawn as needed, almost like a video game. MCP technology. LLMs talk in natural language, so there’s no special formatting requirements.

That’s the nature of the LLM. This flexible structure allows for both flexible input and output. So given a task to do, the LLM can ask, are there any MCP tools around? Then the tools that have registered themselves will announce their existence and their purpose, so the AI can then decide if particular MCP servers can be used in specific instances. Windows has an MCP registration feature, so you can basically include an API in Windows from, let’s say, an enterprise system, and then no additional programming needed. All that’s needed is to give instructions to the LLM to define a specific task for this particular instance.

These moves, subtle as they may seem, actually made tons of AI integrators move into the cloud AI of anthropic, and that’s why now all of the LLMs support the model context protocol. The MCP technology provides two methods of dealing with an LLM. Some MCP servers are called resources, meaning they are sources of information to supply additional context to an AI. In the case of a business use example, an MCP server could provide information into product manuals or procedure manuals, something that is only privately available. Some MCP servers can provide current internet data and many of the chatbots now auto-activate the internet search feature into their chatbots.

For example, Grok is very good at this. But some of the other MCP servers are tools, meaning they will execute the task that an LLM requests and base on the user requirements. So a claims processing agent may be used to make offers to the insured or handle reimbursements. These are action-based. A decision is made and something results from it, in this case, a financial transaction. Quality control of agents. So human intervention and review usually requires higher level employees who are used to screen exceptions and take outliers out of the loop.

Initially, this will be a learning process, so I doubt this would result in immediate layoffs of real employees. However, at some point, some of these projects will yield enough confidence that staff reductions will occur. This is the only reason these agents are going to be used. They’re not used to make the employees happy. This is truly interesting how Microsoft tries to have two separate instances of the OS bifurcate into separate lifelines. The first one that we are so familiar with is the AI companion model where the AI sees what you see, hears what you hear, and lives alongside you.

Famous lines from the Microsoft AI CEO. Personally, I don’t understand why they keep pushing this as the reception of the public is that it is a fail. Nobody wants an AI companion. In fact, Apple and Google don’t actually have AI companions that they both promised. They just intend to provide typical chatbot features and then have fancy things like video editing, language translate on the fly, and things like that. But Windows is the only one super focused on being the AI superspised system currently. So if this is your understanding of AI, yes, good riddance to that.

I think Microsoft will fail here. But wait. But wait. I’m not sure Microsoft cares if that fails. They would likely say nice try, but we’ll be back later. However, the truly scary elements of an independent running computer doing human tasks is where they actually will make money, tons of money. So while I cheer that privacy will win out over Microsoft, unfortunately, Microsoft will potentially end up being more powerful than before, and it’s already the second largest company in the world by capitalization. They’re the biggest investor in OpenAI and all the OpenAI servers are running on Microsoft Azure in the cloud.

Microsoft announced last month at Ignite that they are the agentic OS. In fact, currently they’re the only agentic OS. So while people like me were focused on fighting them on Windows Recall, as it turns out, Windows Recall isn’t even that important in their scheme for global domination. It was really about running independent bots. This really reminds me of the new military scheme where the main leader of a fighter squadron is an actual F-35 with a human pilot, but the other members of the team are endural drones and controlled by AI on the F-35.

That’s the kind of initial model and vision in the enterprise world. Not complete human absence, but a few high-end people controlling a bunch of bots. Does this mean we will see layoffs starting soon? In general, I think enterprises are more cautious and are not necessarily planning on all-out reductions of entry-level jobs. But it is safe to say that new Gen Z graduates in some fields are having a hard time getting jobs. While enterprises are looking for measurable effects of AI on the bottom line, that wasn’t happening in the past year because the technology was not mature enough.

And it was expensive to make autonomous agents, but suddenly this just got easier and I would bet that lots of AI integrators got attracted to Microsoft’s approach and the whole MCP concept. Vibe Coding A heavily impacted area are the new computer science graduates. And that makes sense, less junior programmers are needed since AI can create a lot of the code now and this style of writing code was recently given the name Vibe Coding. Which is actually hilarious to me. When I was young and working as a systems analyst at an insurance company, I would actually spec out an actual full application and then I would hand those specs to the programming team.

So in essence, that spec document I used to write would be the input of an AI in Vibe Coding. All natural language, explanations of inputs and outputs, UI, error conditions, and then the database schema, the transport. Later as a software architect, I would do all that myself together with a team of programmers. But strangely, the way this works now, I probably wouldn’t need a team of programmers like I did in the past. Vibe Coding or the actual capability of AI to write good code only really happened in late 2025.

So this is likely to shine in 2026. What do we do? What do we do in this new environment? If you’re retired or own your own business or are financially independent, then you can just spurn the Microsoft AI companions and all the other AI companions and then let the enterprises worry about themselves. And you go to the beach and relax. However, it is hard to do that because even if you’re already older and are secure in your future, the problem is your children and your children’s children. How do you need to deal with this technology? Well, unfortunately, the answer isn’t to ignore AI or assume that it doesn’t have the capabilities I say here.

2026 is a big turning point. Models don’t have to be more intelligent to take over jobs. They just need effective agents. You need to draw up misconceptions about AI that are irrelevant, for example, constantly connecting AI to politics or to personalities. While that may push you to use one LLM over another because you perceive bias, that is not consequential to business processing. Business processes are defined by specific rules and limited actions. Meaning the only way to fight back is by learning how AI works, learning why AI hallucinations occur, learning why AI doesn’t know who the current president is or anything about assassination attempts.

There’s a reason why AI comes up with these errors, and they will fix that. In the meantime, if you understood the limitations, you wouldn’t have asked those questions to begin with. That fault actually lies on you. Folks, nothing changes here. The concept of an AI companion is crap. It’s an invasion of privacy, period. But while I convinced many of you to turn off Copilot or switch to Linux for your personal news, the reality is that the enterprise world will be totally uninfluenced by my rants. So you all have to know AI well.

I’m a heavy AI user. I used to use Olama, which is a self-hosted AI, but it has limitations. So now I use Grok mostly. I also know its limitations well, but I can’t teach Grok anything directly. It won’t collect my data, though I forced it to do so. Apparently it’s not a privacy threat that way. When I have to do research, it can read manuals faster than I can. I have Tesla on full self-driving. I’m likely on self-driving 80% of the time. Yeah, I heard that snarky comments that the privacy guy uses a self-driving car.

Ha, ha, ha. Funny to those people, but victimhood is not my calling. I learn technology, then I learn to fight it if it goes against my objectives. Final thoughts. I don’t want to create false threats or imagined threats if the direction of the threat is not obvious. This is strange for me to say, but right now the main privacy threat from AI is the Microsoft co-pilot, your AI companion. This is the personal side of things with Windows Recall and to see what you see technology. On the agentic OS side, someone could install other agents on your device that can spy on you and utilize your Windows Recall information.

On the enterprise side, that we cannot control. I can’t help but think Microsoft will still win here, but we have to understand deeply how it works so we can be participants in the infrastructure rather than be the laid off workers. And I have good news for you. I’ve tested this. An LLM cannot replace me, the internet privacy guy. It cannot write script for me. So I’m immune from AI agents. Folks, if you’re serious about privacy, come join us at Braxme. It’s the growing community where real privacy people hang out. No censorship, no nonsense.

While you’re there, check out the tools we actually built and use ourselves. Braxmail. Unlimited aliases, no IP leaks. Brax virtual phone, real anonymous numbers. Bites VPN, no logs, piehole DNS, no big unknown corporation. The Google phones and more in the store. The Brax 3 phone second batch is open for pre-order right now at Braxtech.net, the first batch sold out shortly after release. Big thanks to everyone supporting us on Patreon, Locals, and YouTube memberships. You keep this channel alive. See you next time. [tr:trw].

See more of Rob Braxman Tech on their Public Channel and the MPN Rob Braxman Tech channel.