Architecting Hybrid AI for Infrastructure Operations

Overview

Executive Summary

Enterprises are accelerating AI adoption but quickly discover that no single vendor or model can meet all operational needs. The result is an unavoidable hybrid approach that blends vendor-provided reasoning agents with internally built infrastructure-aware agents.

Making this work safely requires a clear separation between agent reasoning and infrastructure execution. Itential FlowAI, the agentic harness of the Itential Platform, delivers that separation. FlowAgent Builder gives teams the environment where agents reason, plan, and orchestrate, while the Itential Platform’s governed execution engine, the same engine that already runs every workflow and automation on the platform, handles infrastructure actions with strict validation, auditability, and control. FlowMCP Gateway and Itential MCP Server extend that connection to the wider AI ecosystem in both directions. Together, this lets enterprises scale AI across their operational environment without compromising safety, visibility, or trust.

This hybrid model introduces a shared requirement across all customers: the ability to safely attach internal infrastructure to both vendor agents and their own agents without compromising control, observability, or security.

This document outlines the strategy behind Itential FlowAI, now generally available and running in production, and explains how enterprises can adopt AI safely, predictably, and at scale by combining agentic reasoning with deterministic, governed execution.

Market Context

Why Hybrid AI is Emerging in Every Enterprise

The Core Customer Need: Safe Attachment of AI to Infrastructure

Regardless of industry, size, or technical maturity, customers converge on the same realization when exploring AI‑driven automation:

Enterprises are responsible for production systems worth millions or billions of dollars. They cannot risk unbounded actions from autonomous systems, nor can they accept opaque execution models. This leads to several non‑negotiable requirements:

At the highest level, satisfying these demands means separating two things that need opposite properties: a deterministic, instrumented, contract-driven side that all agents, internal or vendor, must use to interact with infrastructure, and a flexible, expressive side where reasoning, planning, and orchestration logic live, routing for human or automated approval.

This separation is the only model that scales safely. In practice, it breaks down further into four operating layers, reasoning, execution, security and governance, and instrumentation, which the rest of this paper walks through in turn.

The Framework

The Itential Hybrid AI Operating Model

Enterprises exploring AI for operations quickly realize they need more than intelligent agents. They need a clear operating model that defines how AI should interact with infrastructure in a safe, predictable, and compliant way. This model is built on three layers that describe how reasoning, execution, and instrumentation work together across hybrid environments, and each layer has a real, named counterpart in the Itential Platform.

AI Reasoning, Delivered by Itential FlowAI

This is where agents interpret intent, evaluate operational state, and generate plans. Agents use enterprise context and learned patterns to think through tasks but do not act on infrastructure directly. FlowAI is the agentic harness that lives in this layer, it is not the whole operating model, just the reasoning piece of it: FlowAgent Builder, FlowAgents, FlowAgent Projects and Sessions, and Model Registry all live here.

Deterministic Execution, Delivered by the Itential Platform

This layer takes agent plans and executes them through validated, contract-aligned actions. It applies schemas, permissions, policies, and full auditability to ensure that every operation is safe and consistent across networks, clouds, and APIs.

Infrastructure Instrumentation, Delivered by Itential Gateway, APIs, & MCP connectivity

This provides the operational data, telemetry, controllers, and automation capabilities that expose your actual network and infrastructure in a controlled and observable way, primarily through Itential Gateway and the Platform’s API and adapter catalog, with FlowMCP Gateway and Itential MCP Server extending that same connectivity specifically to AI systems.

Together, these layers form the Itential Hybrid AI Operating Model. Agents reason, execution stays governed, and infrastructure is safely instrumented and reachable, the foundation for a trusted approach to AI-driven operations. The next sections walk through each layer in that order, starting with reasoning.

Agentic Reasoning

Reasoning Layer: Itential FlowAgent Builder & FlowAgents

FlowAI is the agentic harness of the Itential Platform, the product that lives in the reasoning layer just described. It gives teams a way to build, deploy, and monitor AI agents that reason through goals and act through automation the Platform already runs, generally available today and running in production.

A FlowAgent is an AI agent purpose-built for infrastructure work: given a goal in plain language, it figures out what needs to happen, in what order, using what information, the way an engineer would think through the same problem. FlowAgent Builder is where these agents are created, customized, and governed, giving customers the fastest and safest path to defining agents that understand their infrastructure at a deep operational level.

Once built, a FlowAgent runs a ReAct reasoning loop: it observes current infrastructure state, reasons through the goal, selects the right tool, acts, then observes the result and reasons again. Every action it selects still routes through the execution layer, calling workflows, automations, APIs, and Itential Gateway services as structured tools. The agent plans. It never executes directly.

An agent is assembled from four parts in FlowAgent Builder, all configured before it ever runs:

• Tools
  Every workflow, automation, API, and Itential Gateway service the agent is allowed to call, with structured inputs, defined outputs, and governed execution.
• Skills
  Repeatable task expertise encoded as structured agent instructions, how a network engineer would handle a vulnerability response, an OS upgrade, or a config drift.
• Context
  Live infrastructure state, not stale documentation: real device inventory, real config state, real ticket history.
• Governance
  Who can build the agent, what it can call, and who can execute it, configured at build time and enforced by the Platform at runtime.

FlowAgent Projects group related agents under shared roles, permissions, and export/import portability between environments. FlowAgent Sessions give operators a chronological execution trace for every run, reasoning steps, tool calls, and linked jobs, with the ability to pause, resume, or cancel. Model Registry governs which third-party reasoning models are even available, per environment, and to which builder groups.

When you’re operating infrastructure at Lumen’s scale, the question was never whether AI could help. It was whether we could trust it in production, and Itential’s FlowAI answered that. Our teams were building production-ready agents in minutes, within the same governance and access controls we already rely on.
– Greg Freeman, Vice President, Network and Customer Transformation, Lumen

Governed Execution

Execution Layer: Deterministic Execution on the Itential Platform

Every FlowAgent’s plan hands off to the Itential Platform, the same governed execution engine that already runs every workflow and automation on it today. FlowAI doesn’t introduce a second execution surface for infrastructure teams to learn or trust, it connects a new reasoning layer to the execution layer that’s already proven. This is where a plan actually becomes a change, through a specific set of mechanisms:

Even when an agent generates a novel plan, this engine guarantees that only safe, validated, contract-bound actions are executed. Inputs must conform to schemas, outputs follow known structures, and there are no silent retries or undefined behaviors, whether the trigger is a FlowAgent, a scheduled workflow, or a person.

Trust & Compliance

Security & Governance: How the Platform Keeps Execution Trustworthy

Determinism alone isn’t enough, an execution engine also has to be provably safe to hand control to. Beyond the identity, access, and audit controls above, this also means two-layer agent RBAC, where project-level roles govern who can build and manage agents while agent-level permissions govern who can run a specific one; approval gates and autonomy thresholds, configured per operation type and blast-radius level, with Work Center giving operators one queue for every decision waiting on human input; a secure runtime environment, where FlowAgents run inside the Platform’s governed environment with no way to bypass its guardrails; and zero data retention, meaning the platform keeps no copies of infrastructure data and is SOC 2 Type II by default.

Every control described in the previous section applies identically to a human clicking a button, a scheduled workflow, and a FlowAgent acting on its own. That consistency, not any single control on its own, is what makes it possible to hand an AI agent real operational authority: nothing an agent does is held to a lower standard than what a person or a workflow would be held to.

Connectivity & Reach

Instrumentation Layer: Integrating With Your Actual Network & Infrastructure

Infrastructure Instrumentation is the layer that exposes your real network and infrastructure to reasoning and execution in a controlled, observable way, and it’s a broader story than any single connectivity method.

• Itential Gateway
  The same Gateway that carries out execution also instruments what’s out there: controllers, APIs, scripts, cloud services, and network infrastructure, giving the Platform live, observable access to real devices and services rather than a static inventory.
• APIs, Adapters, & Integrations
  1,000+ existing integrations and adapters, so existing Ansible playbooks, Python scripts, and OpenTofu plans become callable, observable tools without being rebuilt.
• Eventing, Telemetry, & Itential Insights
  Real-time event streams and metrics let FlowAgents reason with live state, and Itential Insights adds step-level execution detail on top, timing, attribution, retry history, so a failure shows exactly what step, what system, and why.

AI-to-AI connectivity is one part of this instrumentation story, not the whole thing, and as the AI ecosystem has matured, it’s become a two-way story of its own.

FlowMCP Gateway: Reaching Out

As infrastructure vendors ship their own MCP servers, FlowMCP Gateway lets FlowAgents call those external tools, such as NetBox, Selector, and Forward Networks, for intelligence enrichment before or during execution, with every external tool inheriting the same RBAC, schema validation, and audit logging as anything built natively.

Itential MCP Server: Reaching In

Itential MCP Server is the inbound counterpart: it connects Claude, ChatGPT, Gemini, or any MCP-compatible model to the Platform’s live execution layer, so external AI systems can query the status of a change, a running workflow, or a device state in plain language, sourced from what’s actually running right now.

Together, Itential Gateway, the API and adapter catalog, and FlowMCP Gateway and Itential MCP Server are what let reasoning and execution actually reach real infrastructure, and the wider AI ecosystem, instead of stopping at a diagram.

Why It Holds Together

Why Reasoning & Execution Require Separate Frameworks

Enterprises often attempt to unify reasoning, orchestration, and execution into a single consolidated framework. On paper, it sounds appealing: one system to think, decide, and act. In practice, it collapses almost immediately under real-world operational pressures, because the needs of reasoning and the needs of execution move in opposite directions. To be clear about which layers this means: “reasoning” is the reasoning layer just covered, and “execution” here means everything downstream of it, execution, security and governance, and instrumentation together, the deterministic side of the Platform.

Why the Deterministic Side Must Stay Rigid

Execution must enforce strict schemas and contracts, produce predictable, testable outcomes, provide complete instrumentation and auditability, and uphold strict security and access boundaries, ensuring that every single action, no matter where it originates, is safe and observable. This side cannot tolerate ambiguity or improvisation.

Why the Reasoning Side Must Stay Flexible

FlowAgent Builder addresses the opposite challenge: empowering customers to create reasoning agents that run directly on the Itential Platform, leveraging Platform-native tools for orchestration, transformation, validation, and integrations. These agents are capable of long-horizon planning, dynamic decision-making, incorporating domain knowledge, handling partial or uncertain input, recovering from unexpected scenarios, and adapting policies and workflows over time. This side must remain expressive.

If unconstrained reasoning is given direct operational authority, outputs become unpredictable, safety boundaries erode, auditing becomes incomplete, and infrastructure becomes vulnerable to malformed commands and untested edge cases. If execution logic is embedded inside the agent layer instead, agents become brittle, every infrastructure change requires rebuilding the agent, reasoning becomes constrained by rigid schemas, and innovation slows. Collapsing both into one layer produces a system simultaneously too unsafe to trust with production infrastructure and too rigid to harness the power of agentic reasoning.

The Separation Creates a Virtuous Cycle

Splitting reasoning and execution into two purpose-built frameworks does more than solve an architectural problem. It creates a positive feedback loop where each layer reinforces the strengths of the other.

• Safety becomes systemic instead of reactive.
  Because reasoning agents cannot bypass the execution boundary, hallucinated commands cannot execute, misinterpreted instructions cannot propagate directly into production, and vendor agents cannot accidentally exceed permissions.
• Agility increases instead of decreases.
  Infrastructure teams can update tools, schemas, credentials, and execution behavior without touching agent logic. AI teams can iterate on reasoning, prompting, or planning strategies without risking execution pathways.
• Flexibility across vendors and models.
  Because FlowMCP Gateway and Itential MCP Server expand the available toolset and connectivity, enterprises can adopt agentic systems from multiple vendors, switch LLM providers when needed, and deploy different models for different agent types. Execution stays consistent no matter who the reasoning provider is, preventing vendor lock-in.
• Observability becomes unified and comprehensive.
  Because execution is centralized, every action is logged in the same format, compliance audits become simpler, and failures can be traced across agents, vendors, and systems.
• Future-proofing at an architectural level.
  Reasoning models evolve at a blistering pace; execution frameworks evolve slowly and deliberately. Decoupling the two lets enterprises upgrade or swap LLMs without rewriting infrastructure code, and lets agents grow more sophisticated without destabilizing execution.
• A flywheel effect emerges.
  Better reasoning leads to more efficient execution plans. Better instrumentation leads to better reasoning. Better agents uncover more opportunities for deterministic tools. Better deterministic tools enable more ambitious agents.

In Practice

Putting It All Together: The Enterprise Hybrid AI Model

How AI Operates Day to Day in the Enterprise

The architectural separation between reasoning and execution is well established. The true value emerges when you look at how enterprises actually operate day to day: multiple vendors, multiple models, legacy systems, inconsistent APIs, distributed teams, and stringent compliance pressures all converge in one operational environment.

1. Vendor Reasoning Agents at the Top of the Stack

Enterprises will continue to adopt best-in-class agentic experiences from multiple vendors, excelling at long-form reasoning, knowledge synthesis, strategy formation, and natural language interfaces for humans. These vendor agents lack the deep infrastructure context or guardrails to safely interact with operational systems, so they connect through Itential MCP Server. Vendor agents, no matter how powerful, remain reasoning-tier systems, not execution-tier systems.

2. Internal Infrastructure Agents Built With FlowAgent Builder

In parallel, enterprises build their own agents using FlowAgent Builder: deeply aware of infrastructure realities, tuned to internal workflows and operational logic, connected to domain knowledge like topology, policy, and compliance, and capable of multi-step procedures across cloud, network, security, and ITSM systems.

3. The Itential Platform as the Universal Safety & Execution Layer

The Itential Platform sits between reasoning and infrastructure, acting as the controlled gateway for all operations. All agents, vendor or internal, use its tools for execution. Every action passes through deterministic schemas and strict validation. Audit logs, metrics, and traces become uniform and centralized. The enterprise retains full control over what infrastructure is exposed and how, no matter which agent initiates the plan.

4. Unified Governance, Policy, & Observability

Because all execution flows through the Platform, governance becomes dramatically simpler: unified policy enforcement, unified access control via RBAC and secrets management, unified audit trails, and unified operational data, metrics, tool performance, error rates, and compliance checks, all feeding a single telemetry plane.

5. A Cohesive, End-to-End AI Operating Model Emerges

Human, Vendor Agent, Internal Agent, Governed Execution and Security, Infrastructure: a clean, layered pipeline where humans engage naturally with reasoning systems, vendor agents provide high-level intelligence, internal agents translate strategy into real operational plans, the Platform’s execution engine and the security controls wrapped around it guarantee safe and controlled action, and infrastructure remains protected at all times.

6. The Hybrid Model Future-Proofs the Organization

This architecture positions enterprises to handle the rapid evolution of vendor LLM offerings, new execution technologies and automation patterns, expanding internal automation at scale, and increasing compliance and security requirements, avoiding lock-in while supporting continuous improvement.

Where This Leads

Conclusion

Enterprises adopting AI must embrace a hybrid operating model if they want both innovation and stability. Adopting a hybrid model without a unifying architectural foundation leads to fragmentation, inconsistent governance, and operational risk. The Itential Hybrid AI Operating Model solves this by keeping reasoning, execution, security, and instrumentation as distinct, purpose-built layers, generally available today, not a roadmap item.

• Reasoning, delivered by Itential FlowAI.
  FlowAgent Builder and FlowAgents give enterprises a platform-native way to define powerful, infrastructure-aware agents capable of reasoning, orchestration, and operational decision-making.
• Execution, delivered by the Itential Platform.
  The same deterministic engine, Workflow Engine, Itential Gateway, APIs, Configuration and Compliance Manager, and Lifecycle Manager, that already runs every workflow and automation today.
• Security and governance, wrapped around every action.
  RBAC, approval gates, audit trails, and secure runtime, applied identically whether a human, a workflow, or a FlowAgent triggered the change.
• Instrumentation, reaching real infrastructure and the wider AI ecosystem.
  Itential Gateway and the API catalog connect to real devices and services; FlowMCP Gateway and Itential MCP Server extend that same reach to AI systems specifically, in both directions.

Together, these layers let enterprises safely attach infrastructure to AI systems, build sophisticated agents directly on the Itential Platform, integrate vendor reasoning agents without exposing infrastructure risk, and maintain consistent governance and auditing across every AI-driven operation.

Independent analysts agree. 451 Research has described FlowAI’s general availability as a strategic repositioning, from network automation vendor to agentic operations platform, naming build-time governance the technical differentiator that addresses AI’s real barrier to production: trust, not capability. Appledore Research calls the approach “AI first, not AI augmented,” a distinction they treat as architectural, not marketing. And Futuriom has named Itential to its Futuriom 50 list six years running.

The Itential Hybrid AI Operating Model gives enterprises the confidence to accelerate AI adoption without compromising trust, control, or operational integrity. It establishes a durable architectural foundation where AI reasoning, governed execution, and infrastructure instrumentation reinforce each other, unlocking a scalable, secure, future-proof model for enterprise automation. For a closer look at how organizations move through this adoption journey stage by stage, read The AI for Infrastructure Journey: From Experimentation to Autonomous Operations guide.