Key Considerations for Architecture, Governance, and Orchestration in Enterprise-Scale DeploymentsRead the full Research Brief below, or click the image above to download a PDF.SummaryAgentic AI is emerging as the next phase of enterprise AI adoption. The shift is subtle but important. Instead of systems that respond to prompts, enterprises are beginning to explore systems that can plan, reason, and take action across business processes. The potential is clear: automating multi-step workflows, accelerating decision-making, and extending agents deeper into operational execution to improve efficiencies.But while the ambition is evident, the path to production is not. Despite widespread investment, most enterprise deployments remain in pilot phases or narrowly scoped use cases. The limiting factor isn’t model capability, but the infrastructure required to support these systems in real-world environments.Agentic systems depend on continuous, governed access to enterprise data and must operate within the same parameters as mission-critical IT for security, governance, performance, and reliability. As these systems scale, architectural cracks begin to surface. Data moves between systems, governance fragments, and security controls become inconsistent. What works in a controlled environment becomes difficult to manage in production at scale.This is driving a broader architectural rethink. Rather than treating AI as a layer applied on top of existing systems, enterprises are starting to evaluate where control should reside — where data access is enforced, where governance is applied, and where execution takes place. As agentic workloads scale, the data layer is emerging as a logical control point.This paper examines that shift, the architectural options available to enterprise IT, and the implications for building and operating agentic systems at scale.Market Overview: The Agentic AI Enterprise Reality CheckThe adoption data tells a clear story about where enterprise agentic AI stands today. According to Deloitte’s 2026 State of AI in the Enterprise report, nearly three-quarters of companies expect to deploy agentic AI within two years, yet only 23% use it even moderately today — and only 21% have a mature governance model in place.That gap is structural, rooted in infrastructure challenges that most current agentic AI frameworks were not built to solve. Moor Insights & Strategy (MI&S) identifies three primary barriers that define the production threshold for most organizations:Data fragmentation. Agents require real-time, governed access to enterprise data to reason and act effectively. Most organizations operate across fragmented data estates — ERP, CRM, data warehouses, SaaS platforms, and object stores — with no unified access layer. Agents operating downstream from that fragmentation produce incomplete or unreliable outputs, while the pipelines required to unify that data introduce complexity that compounds over time.Security at agent scale. Traditional access controls were designed for human users and applications, but what works for human access does not cleanly translate to autonomous systems operating continuously. Agents introduce fundamentally different risks. They can autonomously generate queries, expose sensitive data through prompt injection, and operate at a scale where audit and oversight become difficult. Increasingly, agents can invoke external tools and APIs, extending risk beyond the database itself.Orchestration complexity. Multi-agent systems require assembling LLM APIs, vector stores, RAG pipelines, memory layers, and workflow orchestration into a coherent system. This creates a complex and fragile architecture. Each additional component introduces multiple failure points, governance gaps, and operational overhead. Most IT teams that MI&S encounters are not equipped to build and maintain these systems at scale.The implication is straightforward: Solving for agentic AI in the enterprise requires rethinking where control is established — not at the application layer, where governance is difficult to enforce consistently, but at the data layer, where access, security, and execution can be anchored in infrastructure that enterprises already trust.How Enterprises Are Approaching the Problem — and Where Each Approach Hits a WallThe production gap is not for lack of trying. As enterprises have experimented with agentic AI, several distinct approaches have emerged. Each has genuine strengths, along with inherent structural limitations that become visible as deployments scale.The Build-Your-Own StackThe most common starting point for technically sophisticated organizations is assembly: combining open-source orchestration frameworks with commercial LLM APIs, purpose-built vector databases, custom RAG pipelines, and whatever data connectors the team can integrate. The appeal is real, because it enables maximum control, model flexibility, and the ability to optimize each component independently.In practice, this introduces structural challenges. Data moves between systems — from operational databases to vector stores, and from vector stores to context layers. This leads to synchronization lag and duplication risk. Further, policies governing data access don’t extend across the stack.For startups and technology-native enterprises with strong AI engineering teams, this approach can work. For the broader enterprise market, it becomes difficult to sustain.Figure 1: Build-It-Yourself Agentic AIBuild-your-own agentic stacks are difficult to govern and sustain at scale. Source: Moor Insights & StrategyCloud-Native Orchestration PlatformsCloud providers offer managed orchestration platforms that abstract much of the integration burden and accelerate deployment. These platforms work well when data is already in the cloud, ready for AI use, and governed within the provider’s security model.However, enterprises with on-premises data estates, hybrid environments, or strict data residency requirements find that this approach creates significant friction. This is especially true in regulated industries, where data movement is often the hardest part of the problem to solve.There are also longer-term considerations around vendor lock-in. When orchestration, data, models, and infrastructure converge under one vendor, switching becomes painful, and costs become real.Standalone AI Data InfrastructureA third approach has emerged around purpose-built AI infrastructure — vector databases, embedding services, and specialized agent memory systems layered between operational data and agent frameworks.But purpose-built rarely means purpose-governed. A vector database optimized for fast similarity search was not designed with enterprise access controls, transaction guarantees, or compliance audit trails as primary design objectives. What can accelerate early deployments can also become too complex in the real world.All of these approaches are moving the industry forward, but they share a common limitation. They treat the data layer as something to integrate with, rather than the foundation to build upon.That distinction is at the core of Oracle’s approach.Oracle’s Architectural Response: Anchor Agents in the DatabaseOracle’s approach to the enterprise agentic challenge reflects a clear and differentiated architectural philosophy. While most enterprise agentic AI frameworks are built as orchestration layers that call external databases and services, Oracle has inverted the model by architecting AI capabilities into the database and letting agents execute close to the data they depend on. This can be viewed as a control plane for agentic AI — or an AI operating system that manages critical components right next to the data.The logic builds on Oracle’s existing strengths. Its converged database supports relational, vector, graph, JSON, spatial, and columnar data in a single unified engine. It also includes the security, consistency, transaction guarantees, and operational reliability required for mission-critical enterprise workloads. The Oracle AI Database 26ai release extends this foundation by enabling agent memory, as well as agent creation and execution, within the same system.For regulated industries —or any organization that values data privacy — this matters. Architectures that require data to leave well-governed environments introduce compliance risk that may be unacceptable. Oracle’s model keeps data where it belongs and brings AI to it.From an MI&S perspective, this represents a meaningful differentiation. While Oracle’s model doesn’t totally eliminate complexity, it greatly simplifies AI implementation by locating the architecture in a layer that enterprises already know how to manage. Prioritizing governance and consistency over architectural flexibility, which can become unwieldy, aligns well with the needs of many larger enterprise IT organizations.What follows is a closer look at five capabilities Oracle has introduced to support its agentic AI approach. Taken together, they reinforce a clear point of view: AI should be built where the data already lives, not stitched together across fragmented systems.Private Agent Factory: Agentic AI Built for BusinessOracle AI Database Private Agent Factory is a no-code platform that enables business analysts and domain experts to build, test, deploy, and manage data-centric AI agents and multi-agent workflows. It runs as a container that’s deployable on-premises, in a private cloud, or in any public cloud within the customer’s own tenancy. This keeps enterprise data within the customer’s security perimeter throughout the agent lifecycle.Three pre-built agents are available out of the box: a Database Knowledge Agent for RAG-based information retrieval, a Structured Data Analysis Agent for schema-aware semantic querying, and a Deep Research Agent for multi-step complex analysis. These cover the most common enterprise agentic use cases and enable rapid deployment without custom development. For organizations that need to go further, the Agent Builder enables code-free custom agent design, combining LLM components, data connectors, APIs, and specialized sub-agents into hierarchical, multi-agent systems.Private Agent Factory’s most consequential design decision is not its no-code interface, but its deployment model. Running as a container in the customer’s own environment eliminates the architectural pattern that most enterprise security teams find hardest to accept: enterprise data leaving the governed perimeter to reach external AI services.Equally important is the platform’s commitment to determinism. Agents deployed in production maintain validated, predictable behavior with controls designed to reduce drift from model updates.Current enterprise agent development approaches require significant developer effort, including assembling LLM APIs, vector stores, orchestration frameworks, embedding models, and custom tool integrations into a pipeline. As noted earlier, the result is architecturally fragile — difficult to validate, hard to govern, and prone to behavioral drift in production.From an MI&S standpoint, Private Agent Factory is well aligned with the needs of this segment of the market and is well suited for building agents that interact with a wide range of enterprise data.Autonomous AI Vector Database: The Right Entry PointOracle’s Autonomous AI Vector Database is aimed squarely at developers and data scientists building semantic search, RAG, and agentic applications. On the surface, it checks the expected boxes: support for Python, REST, and SQL, a clean development experience, and a straightforward path from early experimentation into production. As requirements scale, it expands into the full Autonomous AI Database with a single click, without requiring data migration or re-architecture.And this last point is quite compelling, and a significant distinction. Standalone vector databases work well in early stages. They’re easy to deploy and test, and they help teams move quickly. However, they also introduce a second system into the architecture, which is where things often start to break down. Making this setup work requires synchronization pipelines to keep vector data aligned with source systems, separate security and governance models, and increasingly complex cross-database queries. While manageable in pilots, this model becomes more difficult to maintain in production.Oracle’s model avoids that divergence from the start. Vector data and business data operate within the same governed system, simplifying retrieval, maintaining consistent security policies, and ensuring that auditability isn’t an afterthought.Unified Agent Memory Core: Making Agents StatefulAgents without memory aren’t agents — they’re tools. Every interaction starts from scratch, with no persistent context or accumulated learning. While stateless agents can answer questions, they can’t manage complex, multi-step workflows that unfold over time.Oracle Unified Agent Memory Core provides a persistent, governed memory layer for enterprise AI agents, built natively on Oracle AI Database. It organizes memory across three functional layers that map to how agents use context:Short-term working memory captures ongoing conversation snapshots and context cards, stored as vectors for semantic retrieval.Long-term experiential memory preserves generalizations from previous interactions, learned preferences, and procedural instructions, stored as JSON for structured access.Long-term factual memory maintains knowledge graphs and structured business records, queryable through relational and hybrid semantic retrieval.Current enterprise agent memory solutions are fragmented — separate vector stores, graph databases, and document stores, each with its own security model and consistency guarantees. None of these are auditable as a unified system.In what may seem like a recurring theme, Oracle’s approach co-locates agent memory with enterprise data in the same converged database, enabling low-latency reasoning across vector, JSON, graph, relational, text, spatial, and columnar data. Access controls, transaction guarantees, and retention policies that govern business data also govern agent state. This eliminates the need for external AI caches and multiple trips to isolated data sources, accelerating agents’ ability to execute their intended functions.MI&S sees the Unified Agent Memory Core architecture as addressing one of those critical yet often overlooked infrastructure topics. Organizations that get it wrong could spend years retrofitting agent memory onto stacks that were not designed for it.Deep Data Security: Protecting Against the New Threat SurfaceAI agents introduce a security problem that application-layer access controls were never designed to solve. Agents can autonomously generate arbitrary SQL queries, potentially accessing any data that the underlying service account can reach. These “digital workers” can be manipulated through prompt injection to expose sensitive records. They can also bypass application-layer security checks that were designed to constrain human interaction patterns.Consider how most enterprise agent deployments are configured. Agents connect to databases through highly privileged service accounts — broad credentials granted to the infrastructure, with user-level access enforcement left to the application layer. This model was already imperfect for human-facing applications. At agent scale, where hundreds or thousands of autonomous sessions may be running simultaneously, it breaks down entirely. An organization simply can’t trust every application in the chain to enforce the right policy, every time, for every agent interaction. The math doesn’t work, and the risk quickly moves from theoretical to real.Oracle Deep Data Security addresses this by making the database the control plane. Row-, column-, and even cell-level access rules are defined in SQL and enforced at runtime. An agent can generate any query, but the database enforces what that user is allowed to see.The real-world impact is significant. Prompt injection attacks fail at the database, not the application. RAG workflows over vector embeddings are subject to the same access governance as direct SQL queries. Audit trails are generated automatically, and access rules
