jcardena.com › Blog › The Enterprise Agent Platform, Decomposed
184 posts
EN ES

The Enterprise Agent Platform, Decomposed

AI

Strip the branding off an enterprise agent platform and it is six layers: identity, index, model tier, gateway, orchestration, governance. A map you carry between clouds.

The Enterprise Agent Platform, Decomposed

Every few months there is a new enterprise agent platform, and every one of them is presented as a new world with its own vocabulary. It is easy to feel like you have to learn each from scratch. You do not. I have found that if you strip the branding away, they are the same handful of layers, in the same order, doing the same jobs. Learn that shape once and a new platform stops being a new world. It becomes a familiar structure wearing unfamiliar names, and you can find your footing in an afternoon instead of a quarter.

TL;DR: Most enterprise agent platforms decompose into the same six layers. Identity, who is acting. The index, what they are allowed to know. The model tier, how they think. The gateway, what they are allowed to do. Orchestration, how the work moves. And governance, how you stay in control of all of it. That is the map. Once you carry it, every vendor's platform is just their words painted over these six boxes.

AnyPlatformStrip theBrandingThe SameSix LayersA MapYou Carry
Under the logos, every enterprise agent platform is the same shape.

Every platform is the same shape under the branding

The reason the shape repeats is that the problems repeat. Any system where software acts on behalf of a company has to answer the same questions no matter whose logo is on it: who is doing this, what may they see, what may they use, how does the work flow, and who is watching. A platform is just a particular set of answers to those questions, bundled and named. When you see the questions underneath, the bundling stops being intimidating. You are not learning a product. You are checking how this particular product answered the six questions you already know to ask.

Identity: who is acting

The bottom layer is identity, because everything above it depends on knowing who is acting. In an agent platform that who is often not a person but a piece of software acting on its own or for someone, which is the hard new part. This is where the platform decides how agents get their own names, how long their authority lasts, and how you tell an agent acting as itself from one acting on behalf of a human. Get this layer wrong and nothing above it can be trusted, because every permission and every log entry rests on a name that has to mean something.

The index: what they can know

Next is the index, the platform's memory of the company's knowledge, and the gate on it. This is retrieval, but in an enterprise it is inseparable from access: the index has to return only what the person or agent asking is allowed to see. A platform's index layer is really two promises at once, that it can find the right thing and that it will only find it for the right asker. When you evaluate this layer, you are asking both questions together, because an index that leaks and an index that cannot find things are both failures.

Clean gold horizontal strata stacked in navy
Six layers, in the same order, doing the same jobs, whatever the vendor calls them.

The model tier: how they think

Above the index is the model tier, the thinking. Mature platforms are not one model; they are several, chosen by the task. Cheap and fast for the easy majority, strong and expensive for the hard few, and often a private option for the sensitive. This layer is where cost and latency and data residency get decided, by routing each request to the model that fits it. A platform that hands everything to a single model is either overpaying or under-serving; the tier exists precisely so that thinking is matched to the task, not to habit.

The gateway: what they can do

Then the gateway, the layer between the agents and the tools and systems they can act on. This is the single front door where every tool call passes through, so it is where permission, rate, cost, and audit are enforced in one place instead of scattered across every agent. A platform without a real gateway has a mesh no one can govern underneath, no matter how polished the surface. When you look at this layer you are asking one thing: is there a single place that sees and controls what agents actually do, or is reach spread out where nobody can hold it.

Two different structures revealing the same skeleton beneath
Same shape under different names: learn it once, carry it between clouds.

Orchestration and governance: how the work moves, and how you stay in control

The last two layers are how the work moves and how you keep hold of it. Orchestration is the flow: how a request becomes steps, how agents hand off to each other, how loops are bounded so they end. Governance sits over everything: the audit trail, the policies, the evaluation of whether the system is behaving, and the ability to change the rules and see the effect. These two are what separate a demo from a platform. A demo runs once and impresses; a platform runs continuously, observably, and under control. When you assess a new platform, orchestration and governance are where you find out whether anyone can actually operate it, or only launch it.

GOVERNANCEAuditPolicyEvaluationORCHESTRATIONStepsHand-offsBoundedLoopsGATEWAYWhat TheyCan DoMODEL TIERCheapStrongPrivateINDEXWhat TheyCan KnowIDENTITYWho IsActing
The six layers of any enterprise agent platform: identity, index, model tier, gateway, orchestration, governance.

The takeaway

You do not have to relearn agent architecture every time a new platform appears. Underneath the branding it is six layers: identity, index, model tier, gateway, orchestration, governance. Each one answers a question that does not change from vendor to vendor, and each one is a place where a platform can be strong or weak. Carry the map and a new stack becomes an exercise in translation, not a semester of study. That is the quiet advantage of thinking in layers instead of logos: the shape is portable, so your understanding is too.

JC
Juan Cardena
Enterprise AI & Agentic Systems Architect

Enterprise architect with 25 years across web, software, data, and AI. MIT CDAO ’25. Writing on agentic AI in production.