The Agentic Enterprise: Designing IT Architecture for the AI-Powered Era

Updated October 2026

An agentic enterprise is an organization whose IT architecture treats AI agents as first class workers next to people and applications. That takes four new layers on top of the familiar stack, covering agent runtime, shared meaning, models, and orchestration. Open protocols, agent identity, and tracing tie them together. This guide explains each layer, what changed in 2026, and how to sequence the work.

What is an agentic enterprise?

An agentic enterprise is one where AI agents plan and carry out multi step work across systems, with people setting goals and approving high impact actions. The term was popularized by Salesforce Architects, whose reference architecture is the basis for the 11 layer model in this guide.

The shift is real but early. Gartner predicts that 33 percent of enterprise software applications will include agentic AI by 2028, up from under 1 percent in 2024. The same release predicts at least 15 percent of daily work decisions will be made autonomously by 2028.

Why do traditional IT architectures fall short for AI agents?

Traditional architectures assume a person connects the dots between systems. Workflows are fixed in code, data meaning lives in people's heads, and access is granted to humans or to static service accounts. Agents break each of those assumptions because they choose their own next step at run time.

Anthropic's engineering guidance draws the line clearly. Workflows follow predefined code paths, while agents direct their own processes and tool use. That flexibility is the point, but it also means you cannot review every path in advance. The architecture has to supply guardrails, shared context, and traceability instead.

The cost of skipping this work is visible in the numbers. Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. It also estimates that only about 130 of thousands of agentic AI vendors are real, a trend it calls agent washing.

What are the 11 layers of agentic IT architecture?

The Salesforce Architects model has 11 layers. Five are traditional, four are new, and two run across the whole stack. Other vendors draw the boxes differently, but nearly every agentic reference architecture covers the same concerns.

The agentic layer

This layer is where agents live. It handles registration, versioning, memory, tool access, and retirement. Treat agents like any other production workload, with an owner, a change process, and a way to switch them off.

The semantic layer

The semantic layer gives agents the meaning behind data. It holds business definitions, an ontology of entities and relationships, and governed metrics. Without it, two agents can read the same revenue column and compute different answers. The section below covers why this layer matters so much.

The AI and ML layer

This layer centralizes model access, evaluation, and routing so each team does not wire up its own providers. It is where you set which models agents may use for which data classes, and where you run regression tests when a model version changes.

The enterprise orchestration layer

Orchestration coordinates long running work across agents, people, robotic process automation, and legacy systems. It decides when to hand a task to a human, retry a step, or escalate. Our enterprise AI automation platforms comparison reviews the products that compete here.

How do MCP and A2A fit the integration layer?

Two open protocols now carry most agent integration traffic. MCP connects an agent to tools and data, and A2A connects agents to each other. The Agentic AI Foundation hosts both under the Linux Foundation, which lowers lock in risk for enterprise adopters.

Anthropic donated MCP to the foundation in December 2025. At that point it reported more than 10,000 active public MCP servers and 97 million monthly SDK downloads. A2A shipped version 1.0 in March 2026 and joined the foundation in August 2026 with backing from over 150 organizations.

In practice, wrap each system of record as an MCP server with narrow, well named tools. Use A2A when a task must pass to an agent owned by another team or vendor. Our guide to AI agent communication covers message patterns and failure modes in more depth.

Why does the semantic layer decide whether agents can be trusted?

Agents are only as accurate as the meaning they receive. A model that sees raw table names has to guess what "active customer" or "net revenue" means, and it will guess differently each time. A semantic layer turns that guess into a lookup.

The evidence is consistent. In the data.world benchmark, GPT-4 answered enterprise questions correctly 16 percent of the time over raw SQL schemas and 54 percent over a knowledge graph. An April 2026 study found explicit business semantics added 17 to 23 points across Claude Opus 4.7, Sonnet 4.6, and GPT-5.4.

Those definitions are also institutional knowledge, which makes ownership a design choice. Keeping the ontology and metric definitions in open formats on infrastructure you control means agents from any vendor can use them, and you can switch tools without losing them. Galaxy works with teams to build this kind of semantic foundation on the stack they already run.

Our comparison of RAG vs knowledge graph vs semantic layer covers the tradeoffs between these approaches.

How should security and identity work for AI agents?

Every agent needs its own identity, scoped permissions, and an audit trail that names it as the actor. Shared service accounts and borrowed human logins hide who did what and grant far more access than a single task needs.

Identity vendors now support this directly. Microsoft Entra Agent ID gives agents their own accounts, supports both autonomous and delegated access, and logs activity as performed by an agent. In February 2026, NIST's Center for AI Standards and Innovation launched an AI Agent Standards Initiative with agent security and identity research as one of three pillars.

The threat model is also maturing. The OWASP Top 10 for Agentic Applications, released in December 2025, highlights behavior hijacking, tool misuse, and identity and privilege abuse. Our article on data governance for AI agents covers policies and enforcement at the protocol layer.

What does observability look like for agents?

Agent observability means you can replay any task end to end. That includes every model call, tool call, retrieval, and handoff, joined by one trace ID. Without it, a wrong answer or a surprise bill has no explanation.

The OpenTelemetry GenAI semantic conventions define spans, metrics, and events for generative AI clients and MCP. Using them lets agent telemetry flow into the monitoring stack you already operate rather than a separate silo. Add business metrics on top, such as task completion rate, human escalations, and cost per completed task.

Which design principles should guide agentic architecture?

The Salesforce model lists eight principles. The table restates each one as a practical test you can apply in a design review.

The principle most often skipped is human oversight for high impact actions. Define in advance which actions need approval, such as payments, data deletion, or external messages. Encode that list in the orchestration layer instead of relying on prompts.

What is the maturity path to agentic operations?

The Salesforce model defines four stages. Each stage widens what agents may do, so each needs stronger semantics, identity, and observability before you move on.

Many teams start in stages one and two. The jump to stage three is where shared definitions stop being optional, because agents from different departments must agree on what a customer or an order is.

How should an enterprise start building agentic architecture?

Start with one high value process and build only the layers it needs, then reuse them for the next process. This keeps cost tied to a measurable outcome and avoids building a platform nobody uses.

A practical first quarter follows four steps. First, pick a retrieval or single domain use case with a clear owner and metric. Second, expose the needed systems through a small set of MCP tools with scoped agent identities. Third, define the 20 to 50 business terms that use case depends on. Fourth, turn on tracing and cost tracking before launch, not after.

Enterprise architects play a central role in this sequencing. Our guide to the AI model architecture layers for enterprise architects covers the artifacts and governance frameworks they should own.

Frequently asked questions

What is agentic enterprise architecture?

Agentic enterprise architecture is an IT design in which AI agents are first class participants next to people and applications. It adds runtime, semantic, model, and orchestration layers to the usual data, integration, application, and experience layers. Security and operations run across all of them, so agents can act across systems with clear meaning, limits, and audit trails.

What are the 11 layers of the agentic enterprise?

Salesforce Architects describe 11 layers. Five are traditional, namely infrastructure, data, integration, application, and experience. Four are new, namely agentic, semantic, AI and ML, and enterprise orchestration. Two cut across everything, namely security and IT operations. Other vendors draw the stack differently, but most include the same core concerns under different names.

Do I need all 11 layers before deploying agents?

No. Most teams begin with information retrieval agents on top of a governed data layer, basic identity controls, and logging. The semantic layer, shared orchestration, and multi agent coordination grow as use cases prove value. Building every layer up front tends to produce expensive platforms with no users, which is one reason analysts expect many agentic projects to be canceled.

What is the difference between MCP and A2A?

MCP connects an agent to tools, data sources, and applications through a standard server interface. A2A connects one agent to another agent so they can discover each other, delegate tasks, and return results. Both protocols now sit in the Agentic AI Foundation under the Linux Foundation, so many enterprises adopt both rather than choosing one.

Why does a semantic layer matter for AI agents?

Agents act on what data means, not just on what tables contain. A semantic layer or ontology gives them shared definitions for customers, revenue, and products, plus the relationships between them. Published benchmarks show large accuracy gains when models get explicit business meaning instead of raw schemas, which makes this layer a precondition for trustworthy autonomous action.

How should AI agents be authenticated and authorized?

Give each agent its own identity rather than a shared service account or a borrowed human login. Grant it scoped permissions for autonomous work and use delegated tokens when it acts for a user. Identity products such as Microsoft Entra Agent ID now support this pattern directly, with lifecycle controls and logs that mark actions as performed by an agent.

How do you monitor AI agents in production?

Trace every model call, tool call, and handoff with a shared trace ID so you can replay a full task. OpenTelemetry maintains semantic conventions for generative AI clients and MCP, which lets teams send agent telemetry to the tools they already run. Pair traces with business metrics such as task completion, escalations, and cost per task.

What are the main security risks of agentic AI?

OWASP published its Top 10 for Agentic Applications in December 2025. It highlights agent behavior hijacking, tool misuse and exploitation, and identity and privilege abuse among the top risks. The common thread is that agents act, so a manipulated prompt can turn into a real transaction. Least privilege, input isolation, and human approval for high impact actions are the main defenses.

What is the agentic maturity model?

The Salesforce Architects model has four stages. Information retrieval agents answer questions. Single domain agents run simple orchestrated tasks. Multi domain agents coordinate processes across departments. Multi agent, multi domain orchestration lets many agents collaborate across the enterprise. Each stage needs more shared semantics, governance, and observability than the one before it.

Why do agentic AI projects fail?

Gartner predicts over 40 percent of agentic AI projects will be canceled by the end of 2027. It cites escalating costs, unclear business value, and inadequate risk controls. These are architecture problems as much as model problems. Shared identity, observability, and definitions make costs visible, tie agents to measurable outcomes, and put controls in place before agents gain broad access.