A Practical Guide to Building the Brand Brain, Enterprise Brain, and Orchestration Layer for the Agentic Enterprise

The campaign shipped on brand. Same tokens, same type scale, same photography rules. It also quoted a discount that had been retired in Q3, used a claim legal had never approved, and went out to a customer with an open complaint.

Nobody prompted badly. The agent did exactly what its context allowed. The brand guidelines lived in a ninety-page PDF it could not query. The claim library lived in a lawyer’s inbox. The complaint lived in ServiceNow. The budget ceiling lived in a spreadsheet the CMO checks on Fridays.

That is the whole problem with agents at enterprise scale, and it is not a model problem. The moment agents generate at volume, everything you used to keep in documents becomes a runtime constraint. If it is not machine-readable and enforced before the action, it does not exist.

I mapped this in a one-page infographic a few weeks ago: the Agentic Enterprise Experience Stack for a CMO. Eight department agents, a cockpit above them, two transversal brains holding the guardrails, an orchestration layer on top where the customer actually meets the thing. That map named the products. This article does what a map cannot: it tells you how to build the two brains and the orchestration layer yourself.

Orchestration, not decoration. Same thesis as always. Here it gets a bill of materials.

The map, once

Read it top down. The top band is what the customer meets. The bottom two are what keep it safe. Everything in between has a named owner.

Five things to hold onto before we build anything:

1. Brand governs expression. Enterprise governs permission. Two brains, two questions. Violate the first and you look wrong. Violate the second and you are in front of a regulator.
2. Nothing is shared ownership. Every department agent has a mandate, a toolset and a reporting line. If two agents could both own a decision, neither will.
3. The cockpit is an instrument panel, not a pilot. The human CMO decides plan, budget and priority. The cockpit reconciles eight optimistic reports into one narrative and escalates what it cannot decide.
4. Composed at request time, not pre-built. The experience layer assembles the answer, the page and the journey from governed parts. No net-new components at runtime.
5. Both brains are queried before, not audited after. That timing is the difference between a control and a post-mortem.

Buy it, or build it

There is a category of product now for exactly the thing in the middle of this stack: a company brain, agent-native, shared by humans and agents, with permissions that follow the document. Ekory, a Milan startup, is the cleanest example I have seen of the pitch. One workspace instead of a thousand private Claude accounts. Scoped agents (they call them Ekos) per team, customer or process. Designated trusted users who correct the brain in chat and see the fix go live for everyone. A personal agent that inherits exactly the person’s permissions and no more. Per-document visibility, an audit log, EU hosting, MCP in and out.

That is the right shape. If you have four days and a budget, buy something of that shape and skip to the orchestration section.

If you are a design studio, a product team or an enterprise that needs the brain to sit inside its own perimeter, you can build the same shape with parts you already own or can pull from GitHub. The rest of this piece is that build. For every capability there is an open source route and an enterprise route, and most of you will end up mixing them.

The short version of the matrix, in words. Make when the brain has to live inside your perimeter, when you already own the vault or need to (agencies, studios, regulated industries), when you want to own the door, the log and the gate rather than rent them, and when you have one engineer for a month and a designer who cares. Buy when you need it live in days on standard sources, when EU hosting, SSO and per-document visibility are procurement questions rather than yours, when the team will never maintain a repository but will happily correct an answer in chat, and when the platform you already pay for ships the agent runtime.

Either way, the claims library, the do-not-say list and the five policies are yours to write. Nobody sells those.

Here is the whole open build on one page, with the actual parts. Everything below is a walkthrough of this picture.

Build step one: the Cognitive Brand Brain

What the brand is allowed to say, show and sound like. Written so a machine can enforce it, not so a human can admire it.

The vault is the product. Everything else is plumbing you can swap.

1. Sources: find where the truth already lives

You do not write a brand brain from scratch. You collect it. In most organisations the truth is already spread across six places: Figma variables and components, the brand book, the legal-approved claims (usually an email thread), the tone of voice guide, the product taxonomy, and past work with the decisions that shaped it.

Do a one-day audit. For each item write down three things: where it is, who owns it, and whether an agent can read it today. Most rows will end with “PDF” or “in someone’s head”. That list is your backlog.

Behind the paywall: the vault structure and file formats, the six MCP tools with their contracts, the correction loop, the five policies as code, the agent card template, the cockpit report format, the orchestration build with open and enterprise options, the tools inside each agent (Higgsfield, Live Story, Contentsquare and their open equivalents), the small-business version of the whole stack on Grok Bot, the honest failure list, the four-week plan, and the full resource list with repositories.