The rise of AI agents is transforming how enterprises think about automation, integration, and intelligence. But as organizations begin to scale agentic capabilities, a new challenge emerges: How do you choose the right agent technology for the right use case consistently, and at scale?
Without a structured approach, enterprises risk:
- Agents spread across multiple platforms
- Duplicated capabilities
- Inconsistent governance
- Increased cost and complexity
- Proliferation of proof-of-concepts that fail to scale to production
This isn’t just a tooling problem; it’s an enterprise architecture problem. To address this, we first need to rethink how AI agents operate at an enterprise level.
The forgotten layer in an AI agent
Before diving into technology choices, it’s important to reframe how we think about AI agents. Most practitioners are already familiar with the core components of an agent:
- A user-facing interface (eg chat or voice)
- A reasoning engine powered by LLMs
- Access to data for context and grounding
However, what is often overlooked, and critically important in an enterprise context, is the business logic layer. As shown below, AI agents are not just intelligent interfaces. They are layered enterprise systems combining engagement, reasoning, enterprise logic and data.
An AI agent consists of four key layers:
- Engagement (experience layer): This is where intent is captured and responses are delivered across user and system interactions.
- Agentic reasoning (Intelligence layer): This is the “brain” orchestrating planning, decision-making, and tool invocation.
- Business logic (Execution layer): This is where enterprise control lives. This layer executes actions through APIs, workflows, and policies, ensuring outcomes are deterministic, auditable, and aligned to business rules.
- Data (Data and knowledge layer): This provides trusted, governed context to ensure responses are accurate, grounded, and compliant.
Why the business logic layer matters
While the reasoning layer introduces intelligence, it is the business logic layer that retains control. Without it, agents may generate responses, but cannot reliably execute enterprise actions; decisions aren’t consistently aligned to business rules or policies; and outcomes become difficult to govern, audit, or trust.
In reality, even when not explicitly modeled, this layer is always present, whether as APIs and services, workflow engines, domain applications, or platform-native capabilities. This is where determinism is enforced, transactions are executed, and enterprise policies are applied.
Bringing it together
What’s important is that agents combine probabilistic reasoning with deterministic enterprise capabilities, even when these are abstracted within the same platform. This is true even for agents engineered directly on data platforms. While LLM reasoning remains probabilistic, the data queries, pipelines, and transformations are deterministic. So this distinction still exists conceptually, even if it is abstracted within a single platform.
While reasoning may occur within a single context, enterprise actions rarely do. The business logic layer, where APIs, workflows, and policies are executed, often spans multiple systems and domains. This means that even a seemingly domain-specific agent will frequently need to:
- Retrieve data from other domains
- Invoke external services
- Coordinate with other agents or enterprise capabilities
This is the inflection point because once agents begin interacting across domains, you are no longer building isolated capabilities but are designing an enterprise agent ecosystem.
Agents are a domain problem, not just a tech problem
Let’s take a step back from a single agent to the enterprise. One of the biggest misconceptions is treating AI agents as standalone capabilities. In reality, agents often need to operate across multiple domains across an enterprise (eg customer, finance, supply chain, R&D, HR, data and analytics), as shown below.
A domain-based multi-platform architecture is one where:
- Agents are aligned to business domains
- Agents are built close to the data and business logic that they need
- Domain platforms provide native agent capabilities
- Enterprise services provide governance, orchestration and control
To make this more concrete, consider a sales agent which sits within the customer domain. To make this agent truly effective it cannot operate in isolation.
For example, when responding to a customer inquiry, the sales agent may need to:
- Retrieve real-time product availability from the supply chain and operations domain
- Check pricing plus
- Discounts and promotions from the finance domain
- Trigger order creation or fulfillment workflows across downstream systems
In the following scenario, the sales agent is anchored in a single domain but the business processes it supports naturally spans multiple domains. This is why even domain-aligned agents inherently become cross-domain in execution. However, this is not just a scaling challenge but is a direct consequence of how AI agents are constructed.
The challenge with the architecture outlined here is that building complex agent orchestration logic within the customer domain creates fragmentation, inconsistency and governance risk. The enterprise impact of this is that:
- Agent orchestration logic becomes duplicated across every domain
- Agents become tightly coupled, eroding domain independence
- No consistent enforcement of policy, security, or audit
- Agent integration and workflow sprawl accelerates exponentially
- Enterprise-scale governance becomes unmanageable
The wider result of this can be significant as:
This approach slows down the innovation and benefits of agentic transformation while creating significant risks of loss of control, security vulnerabilities, and cost overrun.
Scaling agents across the enterprise without losing control
Most enterprise IT execs and architect teams successfully build domain agents. What they struggle with is managing how those agents are discovered, interact, scale, and evolve across the enterprise. This is precisely where Agent Fabric becomes critical. This is the architectural layer required to scale AI agents safely and consistently across the enterprise.
As outlined above, Agent Fabric provides the enterprise agent control plane, enabling:
- Agent Discoverability: A central catalog of all agentic assets to enable discovery and reuse independent of where they are built. This includes an Agent Scanner capability, which provides the ability to automatically discover and catalog agents across AWS Bedrock, Google Vertex AI, Azure Copilot and GoDaddy
- Agent Governance and Security: Consistent policy enforcement (eg PII detection, compliance rules, etc.) across all agent interactions with MCP tools and other agents. This includes the ability to manage application access to strategic LLMs while ensuring security, SLAs and governance at scale
- Agent Orchestration (Agent Broker): Context-aware routing service to dynamically select and coordinate the right agents and tools to fulfill a business process. The Agent Broker capability is evolving to a graph-based deterministic workflow, providing the ability to reason and route tasks across agents and tools to automate complex business processes with higher predictability
- Agent Observability: End-to-end visibility into how agents and MCP tools interact both within and across domains
But more importantly, it introduces a new architectural concept: Agent Fabric (ie an Agent Control Plane) is to agentic transformation what API-led connectivity is to digital transformation.
Going back to our earlier example of a customer domain sales agent directly interacting with other agents and MCP tools across other enterprise domains, let’s see how this changes with the introduction of Agent Fabric to handle the cross-domain orchestration using an order management Agent Broker.
In the figure above, the customer sales agent calls the order management agent broker which retrieves real-time product availability from the supply chain and operations domain; it then checks pricing, discounts, and promotions from the finance domain, and lastly, it triggers order creation or fulfillment workflows.
The benefits of this approach are numerous:
- Controlled cross-domain execution: Instead of uncontrolled agent sprawl, cross-domain interactions are explicitly orchestrated while dependencies are visible and governed
- Decoupled agent-to-agent interaction: Agents no longer call each other directly but instead multi-domain agentic interactions are mediated through the broker and domains remain loosely coupled
- Reuse and composability: Instead of duplicating capabilities, existing agents (eg pricing, promotions etc) are reused and composed into a higher-level business process
- Context-aware orchestration: The agent broker understands the intent, dynamically selects the right agents, tools, and services and can combine deterministic rules with AI-driven decisions
- Centralized governance and policy enforcement: All interactions flow through a single control point providing PII detection, compliance, access control policies, etc.
- Observability and traceability: The broker provides end-to-end visibility of agent interactions, traceable decision chains, debugging and audit
Agent Broker transforms agent interactions from ad-hoc, point-to-point calls into governed, observable, and orchestrated enterprise workflows.
Agent Fabric: The control plane organizations need
AI agents don’t fail because of lack of intelligence; they fail because of a lack of control. As agents evolve from isolated capabilities to enterprise-wide systems, the need for an enterprise-level control plane becomes unavoidable. Agent Fabric provides that control plane and enables organizations to move from fragmented experimentation to a connected, governed, and scalable agent ecosystem.
