The five layers of the new operating system of organisations

November 18, 2026

The five layers of the new operating system of organisations

Most conversations about AI inside companies start with the same question: which model should we pick? It is a fair question, and it is asked on the wrong floor. The model is one layer out of five, the only one you replace every six months.

Every industrial revolution automated something: strength, then energy, then calculation. This time, it is intelligence. And intelligence is the engine of all value creation. An organisation that truly takes it in therefore does far more than add a tool to its kit: it changes operating system, that discreet machinery which makes everything else hold together. Before using it, you need to learn to read it: five stacked layers, from data to humans.

Let us follow Claire, a recruiter, all the way up the stack. Monday, 9 a.m., a request lands on her desk: find a project manager. Ten steps, weeks of lead time, almost all of it by hand.

Layer 1: data

Diagram: the first layer of the stack, data, dots arranged in a grid on a slab

At ground level sits the raw material: everything the company produces as it works, about its customers, its operations and its exchanges. For Claire, that means the applications received, the past interviews and the job descriptions. This layer already exists in every organisation, but it is scattered across inboxes, personal folders and software that does not talk to each other.

Yet an AI can only work with what it can read. The first job is therefore the least glamorous of all, since it consists in gathering that data and opening access to it. Once readable, it becomes the base everything else stands on. Claire now has at her fingertips everything the company knows about the role.

Layer 2: the Second Brain

Diagram: the second layer, the Second Brain, a network of linked notes laid over the data

This is the new layer in the model. The Second Brain is the company’s organised memory: what it knows, what it has decided, what it has learned. Where data records what happened, it keeps what was understood from it, as linked notes that teams and AIs alike can read.

Its first job is to give AIs context. Without memory, an AI starting a task produces generic work, whereas once it is fed the history of a file and the decisions already made, it gets things right.

Two choices make it viable. The first is a division of labour, in which the human curates and decides while the machine keeps the records. Classic note-taking methods stumble right there, because the cost of upkeep ends up killing the tool. The second concerns the medium: plain, versioned text files rather than a proprietary database. In the final result, the quality of those files weighs more than the level of the model reading them. And the day you change AI, you plug your memory back in without losing anything. I practise this every day: my own second brain holds about 1,700 linked notes, and every conversation I have with an AI starts there.

For Claire, the difference is immediate, because the Second Brain remembers that the last two project managers left within a year, and why. Her brief starts from there.

Layer 3: interchangeable AIs

Diagram: the third layer, three AI blocks on their bases, only one of them active

Next come the engines: the models and the agents, the ones that do the work. For Claire, the ad gets written, the applications get read, the meetings fall into place.

Their main feature is that they can be replaced. They change every six months and the best ones now sit within a whisker of each other, which lets you compare them and switch without drama. Building your strategy on one specific model therefore means laying your foundations on the layer that moves fastest. This layer also has a limit: a generic AI, however brilliant, knows nothing of your rules or your trade.

Layer 4: business harnesses

Diagram: the fourth layer, business harnesses, three yellow frames marked with checkpoints

A harness is a trade’s know-how put into a form AI can use, with its rules, its memory, its tools and its guardrails. There is one per trade. It is what turns a generic AI into a reliable colleague. The recruiting harness, for instance, rules out any discriminatory criterion and has Claire review every message before it goes out: the human checkpoint is written into the harness itself, from its design.

The weight of this layer can be measured. In March 2026, the LangChain team took its coding agent from 52.8% to 66.5% on the Terminal Bench 2.0 benchmark without touching the model, only by reworking the harness.1 On SWE-bench Pro, changing the harness moves the score by more than 22 points, when switching to the best model of the moment moves it by about one.2 Hence an equation, detailed in my article on harness engineering: Agent = Model + Harness².

Layer 5: humans

Diagram: the full stack, with the human layer on top, five figures in a circle

At the top of the stack stand the humans, as orchestrators: they are the ones who set the intent, decide, take responsibility and carry the relationship.

The split with the machine is made task by task. On some, the machine does the work and the human checks it: we automate. On others, the human stays at the centre and AI multiplies what they do: we augment. Both modes will coexist, and it is up to you to choose where.

Four skills then gain value. First comes intention, meaning knowing what you want and saying it clearly, then judgement, to assess what the machine produces, spot the error and decide. Next come relationships, which means listening, persuading and working as a team, and finally imagination, the ability to see what does not exist yet.

What is left to Claire is what matters: rethinking the role, framing the need, leading the interviews and deciding. In this scenario, what used to take weeks takes a few days, and she spends most of her time with the candidates.

Where the value sits

Once the five layers are in place, the operating system runs, and it reads in both directions: intent comes down floor by floor, then work goes back up, checked at every step.

The remaining question is what, in all of this, truly belongs to you. Models can be bought, and your competitors buy the same ones. Two layers, on the other hand, are in no catalogue: your memory and your know-how, that is, the Second Brain and the harnesses. That is where the value sits: the two yellow layers in the diagram. They have one thing in common, which is that they get built patiently, by the people who know the trade. And that holds all the way to the top. I keep saying it to executives: within less than five years, anyone who is not competent in AI will no longer have a seat on an executive committee.

Not every organisation will move at the same pace. A services company can already run most of its activity through these five layers, whereas a machine manufacturer will have to wait for robotics to take it there. And when this stack runs in enough companies, their systems will talk to each other directly, from one company to the next, without waiting for a human to relay the message.

Where to start

A system like this is intimidating, because everything in it depends on everything else, and that is also what makes it powerful. You still have to start somewhere, and the right starting point is small: one trade, one task.

Pick something you master and still do by hand. Do it once with AI, explaining your rules and your reasons, then ask it to turn that into a reusable procedure, what is called a “skill”. Then test it in a fresh conversation and correct it until it holds. When the second and the third exist, look for the links between them. You will have laid the first stone of a harness. You can start on Monday.

Learning to build your stack

Building your memory and your harnesses is something you learn. That is what Hypervibe is for.

Hypervibe is first of all a two-day training course, designed for people who have never written a line of code. You learn to create your own tools with AI, to give it a memory from one session to the next, to write your first skills and routines, and to connect agents to your tools. It also runs in-house, for a whole team. Discover the Hypervibe training

Two plugins extend the training, and each one matches a layer. Hypervibe Brain installs a second brain: your ideas, your projects and the people who matter, as linked notes that Claude reads in every conversation. The Hypervibe plugin, open source and free, is a complete harness for one specific trade, that of building a web application and putting it online. It illustrates layer 3 along the way, since, designed for Claude Code, it also plugs into Codex, OpenCode or Antigravity.

In six months, you will have changed model. What you have written by then, memory and know-how, will still be yours: that is your operating system.

And in your organisation? This model is also presented on stage, with real examples: AI and the future of work for teams, AI and strategy for executive committees.

Footnotes

  1. LangChain, “Improving Deep Agents with harness engineering” (March 2026). With the model frozen (GPT-5.2-Codex), the agent goes from 52.8% to 66.5% on Terminal Bench 2.0, from the top 30 to the top 5. https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering ↩

  2. “Agent Scaffolding Beats Model Upgrades” (Particula, 2026). On SWE-bench Pro, the scaffold accounts for a gap of more than 22 points, where switching to the best model of the moment moves it by about one. https://particula.tech/blog/agent-scaffolding-beats-model-upgrades-swe-bench ↩

The five layers of the new operating system of organisations | Flavien Chervet