Building our AI Operating System

The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.

Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:

• 7 high-impact AI strategies to accelerate your marketing performance

• Practical use cases for content creation, lead gen, and personalization

• Expert insights into how top marketers are using AI today

• A framework to evaluate and implement AI tools efficiently

Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.

Download the Free Report

FUTURE
Building Our AI Operating System

Drop everything you’re doing because this is the most important post I’ve written this year, potentially since I started this newsletter.

Today, you are going to “feel the AI” in a way you’ve probably never had by putting the first stones in your first true AI personal operating system, one that lets you automate your life one English word at a time.

That is, a virtual computer sitting on top of your own computer, one you’ll control using your preferred natural language.

Among other things, today you’ll learn:

1. Why the time is now to start building your operating system,

2. Why this is not just another workflow automation builder, but a truly agentic system where the AI is in full command. If you need something, instead of building it, just ask for it.

3. What is the digital architecture of the future, including all the key components. You will build:

   1. An email management tool that allow your AI operating system agent to read, write, and move emails all via plain English instructions (i.e., read my email inbox, order it, and send me a voice summary)

   2. A simple trick to create a single source of memory for all models you use, so that, independently of what AI you’re using, from OpenAI to Claude, it knows the key things it has to know about you.

   3. Your first agent workforce. That is, you will learn how you can create out-of-the-box agents just using the power of the English language. In the AI age, everyone has to be an AI manager, and you can train and update them without any technical expertise.

From here on, you put your limits on what you can make these agents do for you. I know, that’s a lot of hype for someone who claims to fight the hype. But that’s the point, for once, I’m that hyped about something.

Let’s build our first AI-led operating system.

Why the time is now

I’ve been calling for the arrival of agents this entire year. They are coming, right? The reality is that what most people, including myself, had not yet realized is that they are here already.

It’s not a coding tool, it’s a supercharged agent. An operating system.

‘Agents are coming’ was something I already told you when OpenAI released the Deep Research API to access their deep research agents, calling them the “best agents on the planet.”

I was not wrong, but I mistakenly assumed I would have to build the entire agentic scaffold on top of it. I thought we finally had the AI, but I also thought I would have to build the agentic behavior on top of it.

As it turns out, that scaffold already existed, and I had not realized because Anthropic made the marketing mistake of calling the best agent product on the planet ‘Claude Code’ instead of ‘Claude Agent,’ or, better, ‘Claude Operating System.’

But why Claude Code? Why is this tool the product that could help you automate everything?

Terrible Naming, Absurd Power

AI companies suck at naming things, and this could easily go down as the worst naming decision. Anthropic really should fire its marketing team, because the name ‘Claude Code’ doesn’t do justice to the immense power this tool has.

But what is Claude Code?

In simple terms, it is a command-line interface, or CLI-based tool, that allows you to connect with AI models using your computer’s terminal. Sounds geeky, you almost feel like an 80s hacker at first, but trust me, it looks way scarier than it actually is.

Since using it as an operating system, I have yet to write a single line of code.

But why is Claude Code so powerful?

When OpenAI released the Deep Research API, I mentioned it was the most powerful model on the planet because it could chain multiple tools in one single chain of thought. Then Kimi K2 came out, only to confirm our intuitions; extensive tool calling was the key agentic behavior.

Yet, while these are indeed great models for tool calling (it is rumored that another Chinese model, GLM-4.5, is the best tool caller on the planet), agents really shine when models with great potential are used correctly.

And the answer to which product achieves that best is Claude Code. But to understand its power, let’s dive deeper into what agent scaffolds like Claude Code are by diving into the architecture we are going to use today.

An AI Operating System

To reiterate my goal for today, I aim to demonstrate that you now possess the necessary tools to develop an operating system powered by AI. But to understand this, let’s first recall what an AI agent is.

Months ago, I went the extra mile to describe ‘what an agent was.’ It’s here in case you want to revisit, but the bottom line is simple.

Revisiting Key Concepts

Three things form an agent system: AI Model (the thinking engine); Memory, components that provide context to the model; and tools that empower the agent to take action. Insofar as an AI model can plan what to do, it can only get as far as what the tools allow.

We can simplify things further by separating the AI model from everything else, and call everything else the ‘scaffolding’ (i.e., everything we add on top of a model to make it an agent).

The question of how much scaffolding an AI model needs is a matter of intense debate. Personally, I am in Noam Brown’s, OpenAI Reasoning Lead, camp: scaffolding must be minimal as more powerful future models will require less scaffolding for every single new iteration.

Put another way, the more control and percentage over the total the AI model represents in the system, the better your system is, and thus the more truthful to the term ‘agentic’ it is.

Consequently, even though you already have agentic products like the general-purpose ManusAI or Genspark, and application-specific ones like Shortcut for Excel, it’s unlikely these tools will be better than the scaffolds provided by the model Labs themselves.

Therefore, you can basically see every single AI startup that isn’t hardware (NVIDIA), infrastructure (Hyperscalers and LLM providers like Fireworks), or Model (OpenAI) companies as ‘scaffold builders’ more generally known as ‘AI wrappers.’

Naturally, this means Claude Code is a wrapper too, but one that has special properties:

1. Wrapped around models created by the same company that created it, ensuring maximum adherence; it’s hard to fathom a future in which any of the AI startups building agents get better results on Claude than Anthropic itself.

2. Best-practice oriented. The model is used as intended. Things like telling the model ‘to make a plan first,’ which is a basic requirement for success with agents, come out of the box (in fact, it’s a tool, more on that later). Also, when creating agents, the product creates the agents automatically using Anthropic’s prompting recommendations.

3. Minimalistic yet ample support for model configurations (behaviors, what model to choose), permissions (what each agent can or can’t do), and cost management (you can see live what you’re spending),

4. Hooks support (certain events can kickstart certain processes, increasing automation coverage)

5. Multi-agent support. Models can saturate context windows fast, so you can break down your agent into multiple agents, each a new instance of a Claude model, also guaranteeing separation of concerns (each agent specializes in what matters).

6. Headless automation. Just released literally 2 days ago, Claude Code now has an SDK that allows users to configure it to run autonomously. For instance, you can run Claude Code by calling it from other applications, such as executing a specific email agent whenever you receive an email, eliminating the need for me to request it explicitly.

In a nutshell, you’re extremely close to the model (no fancy abstractions like other agent frameworks that simplify things but reduce your control over the model), yet surprisingly, with almost no technicalities involved because the tool itself is configured using English.

Let me explain.

Agent-driven Operating Systems are Declarative, not Imperative.

I’ve been rambling about the idea that AI makes software declarative for a long time.

Instead of the traditional, imperative programming paradigm, where if you wanted a machine to do something, you had to explicitly define what had to be done, with AI, the human’s role is less about defining steps and more about setting goals.

But what is this ‘AI operating system’?

An AI operating system is nothing but a virtual computer sitting inside your physical computer that can understand natural language commands, plan how to achieve those demands, gather context, and leverage tools to achieve the goal, executing the plan step-by-step.

In that paradigm, as all logic is absorbed by the AI (with prices going to zero, thereby decreasing overall COGS), and tools take care of the non-logic stuff (logging and sending stuff, using third-party software, maths, physics simulation, etc.), this form factor is just superior for several reasons:

1. By outsourcing logic to the AI, your software becomes more flexible. As long as your tools and context allow for it, you can ask basically anything, with zero added configuration time because the AI will adapt to your request.

2. It’s already cheaper, and will only get cheaper over time as both energy and token costs decrease by orders of magnitude every year. This form factor minimizes human participation, reducing operating costs while overall COGS decrease as model prices plummet.

How can humans compete with a technology that, while imitating us, “cuts its wage” by 90% every year?

So, to summarise it all, I initially thought models were ready, but not the agent tooling.

I was wrong, Claude Code and other initially-conceived-for-code tools are amazing agent scaffolds, the basic pillar of the AI operating system, and I’m going to show it to you today by defining the following operating system:

And remember, the crucial thing to understand is that we are building an operating system without writing a single line of code.

Become a paying subscriber of Full Premium package to get access to this post and other subscriber-only content.

Already a paying subscriber? Sign In.

A subscription gets you:

• • NO ADS
• • An additional insights email on Tuesdays
• • Gain access to TheWhiteBox's knowledge base to access four times more content than the free version on markets, cutting-edge research, company deep dives, AI engineering tips, & more