Skip to content

Architecture

Contents

Overview

Seven layers. The bottom one never moves.

Super Brain describes how an Industry AI Operating System is put together. It is vendor neutral, model neutral, and open.

Read it top down if you want to understand what people experience. Read it bottom up if you want to understand what it touches.

The stack

The seven layers.

Where people meet the system. A question asked, an answer returned, a decision approved.

Specialists, not one assistant. Each has a role, a scope, and a declared limit.

The step between knowing and doing. Reasoning happens in the industry's own terms, not in generic prose.

The layer that makes this yours rather than generic. Industry knowledge is modelled as a graph. Company knowledge stays private.

The reasoning engine. Replaceable by design, and it should be.

How intelligence becomes something that happened, under declared permissions.

Unchanged. That is the whole point. Nothing is migrated, nothing is replaced, nothing is switched off.

Layer 07, human experience

Where people meet the system.

A question asked, an answer returned, a decision approved.

Five distinct surfaces, because five roles need different things. An executive needs a position. A manager needs an exception. An employee needs the next action. A partner needs a boundary. A customer needs an outcome.

The OS routes the same underlying reasoning to each of them at the right resolution.

Layer 06, industry agents

Specialists, not one assistant.

Each has a role, a scope, and a declared limit.

Every agent publishes a manifest stating what it can read, what it can reason about, what it can recommend, what it can execute, and what it can never do without a human.

Agents are composable. A brand question may call the commercial agent, which calls the supply agent. The manifest chain is auditable end to end.

Layer 05, industry reasoning

The step between knowing and doing.

Six operations, in increasing commitment: understand, reason, predict, recommend, plan, simulate.

Simulation matters more than it looks. A recommendation the business can stress test is one the business will act on. A recommendation it cannot is one it will argue with.

Layer 04, knowledge and memory

The layer that makes this yours rather than generic.

Two halves. Company knowledge holds policies, SOPs, products, customers, contracts, documents and the decisions the organisation has already made. Industry knowledge holds the ontology, regulations, terminology, workflows, benchmarks and domain relationships of the sector.

This is a knowledge graph, not a document store. Entities and relationships are modelled explicitly, so the system can traverse them rather than guess at them.

Layer 04 is where an Industry OS stops being a chatbot.

Layer 03, model layer

The reasoning engine. Replaceable by design.

Claude, OpenAI, Gemini, Llama and whatever follows them. Foundation capability improves every few months and gets cheaper. Nothing above this layer should be rewritten when it does.

Model choice is a deployment decision, made per workload, per tenant, and per data boundary.

Layer 02, enterprise action

How intelligence becomes something that happened.

APIs, MCP, workflow engines, RPA where an API does not exist, and the event bus.

Every action carries the identity of the agent that took it, the human who approved it if approval was required, and the reasoning that produced it. An action without a trace is a defect, not a feature.

Layer 01, systems of record

Unchanged. That is the whole point.

SAP, Salesforce, Oracle, Veeva, ServiceNow, Epic, Workday, Snowflake, Databricks, data lakes and the internal systems nobody has documented in years.

The OS reads from them and acts through them. It does not migrate them, mirror them as a permanent copy, or become the new place where the record lives.

Don't rip and replace. Reason and act above.

The three kinds of knowledge

An agent needs three different brains.

Foundation intelligence
Language, reasoning, code, planning, general world knowledge. Supplied by the model layer. Getting cheaper every year.
Industry intelligence
The operating reality of a sector, modelled properly. Entities, relationships, rules, workflows, metrics, decisions, events and actions. Supplied by the ontology and the Industry Pack.
Company intelligence
Customers, products, processes, territories, contracts, policies, SOPs and historical decisions. Supplied by the enterprise, and never leaving it.

The three together produce company specific intelligence. That is the point where an agent stops being impressive and starts being useful.

The formula

Foundation Model

Industry Pack

Company Pack

Agent Role

Enterprise Agent

What the OS provides

Twelve services, none optional.

  • Identity
  • Memory
  • Industry semantics
  • Company context
  • Permissions
  • Tools
  • Workflows
  • Governance
  • Observability
  • Human approval
  • Data access
  • Action access

Plate data

Spec
Super Brain
Version
1.0
Layers
07
Industry OS
05
Hub
Live

Vendor neutral, model neutral, and open.

READ NEXT · 12

Read next.

The Super Brain spec
What an implementation of these layers must provide.
Packs
What fills layer 04, and who owns each half.
Industry Operating Systems
The same seven layers across five industries.
Agent Manifest
How layer 06 declares its limits.
Thesis
Why this layer exists at all, argued in full.

One architecture. Twenty pages. Same document.

TALK TO US · 13

Bring your architecture. We will map it against this one.

Most conversations start with a diagram of what you already run. That is the right place to begin.

Talk to us

A conversation, not a demo script. We ask what your systems of record are, what they cannot answer, and which decisions actually matter. If an Industry OS is not the right answer for you, we will say so.