Beyond the LLM: Engineering Secure, Reliable AI Agents for Enterprise Infrastructure

Share
Beyond the LLM: Engineering Secure, Reliable AI Agents for Enterprise Infrastructure
Photo by fabio / Unsplash

Building a capable artificial intelligence model is one challenge. Integrating that intelligence into business-critical infrastructure, while maintaining security, reliability and operational control, is an entirely different engineering problem.

The rapid advancement of large language models has created considerable enthusiasm around the possibility of intelligent automation. Modern models can interpret complex technical documentation, analyse software, explain network protocols and assist with problems that previously required substantial human expertise. For infrastructure providers and enterprise IT departments, these capabilities offer the prospect of reducing repetitive operational work and improving access to technical knowledge.

However, there is an important distinction between an AI system that understands how infrastructure operates and one that can safely interact with that infrastructure.

A language model may be capable of generating a perfectly valid command to modify a routing configuration, restart a virtual machine or alter a firewall policy. That does not mean it should possess unrestricted authority to perform those actions. Nor does it guarantee that the command is appropriate for the environment in which it will be executed.

This distinction becomes particularly important in multi-tenant cloud environments, where a single infrastructure platform may support numerous customers with different security requirements, service agreements and operational boundaries.

Developing production-ready enterprise AI therefore requires considerably more than selecting a model and connecting it to a collection of APIs. It demands a carefully designed architecture that separates reasoning from execution, applies security controls independently of model behaviour and provides sufficient observability to understand what the system is doing.

Intelligence and Authority Must Remain Separate

One of the most important architectural principles in enterprise AI is the separation of the reasoning component from the systems responsible for enforcing operational authority.

A language model should be treated as a component capable of interpreting information, generating recommendations and proposing actions. The surrounding application must determine which information the model is permitted to access, which tools are available and whether any proposed action can be executed.

This distinction is similar to established approaches in secure software architecture. An application may request access to a protected resource, but the operating system or authorisation service determines whether that request is permitted. The application cannot simply declare that it has sufficient privileges.

Enterprise AI should follow the same principle. A model might request information about a virtual machine, but the integration layer must verify that the requesting user is authorised to access that resource. If the model proposes modifying a production configuration, a separate policy enforcement mechanism should determine whether the operation is allowed and whether human approval is required.

The resulting architecture typically consists of several cooperating components. A user-facing application receives requests and establishes the authenticated identity of the requester. An orchestration service manages tasks and coordinates the necessary operations. One or more models provide reasoning and analytical capabilities, while controlled integration services connect the platform to infrastructure and business applications.

This arrangement is often described as an AI harness or agent execution environment. Its value lies in making the model part of a governed system rather than allowing the model itself to become the authority responsible for deciding what is safe.

Designing an AI Control Plane

A practical enterprise AI platform benefits from a clearly defined control plane responsible for managing execution, permissions, task state and operational policies.

The control plane receives a request, determines the relevant context and identifies which capabilities are required. It can then select an appropriate model, provide access to authorised tools and monitor the resulting workflow.

Consider an engineer asking an AI platform to investigate intermittent connectivity between two data centres. The request may require information from network monitoring, routing configurations, virtualisation infrastructure and historical incident records. Rather than granting the model direct access to every system, the control plane coordinates narrowly scoped requests through approved integration services.

A monitoring service might provide interface utilisation and packet-loss measurements. A network configuration service could retrieve relevant routing information, while a documentation service supplies the intended network topology and operational standards. The model can then analyse the combined evidence without needing unrestricted credentials for the underlying systems.

The control plane also provides a natural location for enforcing execution limits, recording activity and managing approval workflows. If an investigation requires a potentially disruptive action, such as clearing a routing session or restarting a network service, the system can suspend execution until the appropriate authorisation has been obtained.

This architecture becomes especially valuable when organisations operate multiple models. A smaller local model might be sufficient for classifying monitoring alerts, while a more capable reasoning model could be used for complex incident analysis. The orchestration layer can make those selections according to task requirements without exposing the underlying complexity to the user.

The objective is to create a platform in which intelligence can evolve independently of the security and operational mechanisms that govern its use.

The Role of Specialised AI Agents

Although a single general-purpose model can undertake a wide range of activities, there are situations where dividing responsibilities between specialised agents offers practical advantages.

An infrastructure operations platform might employ separate agents for network analysis, virtualisation management, monitoring, security assessment and technical reporting. Each agent would have access to a defined collection of tools and operational knowledge relevant to its responsibilities.

During an incident investigation, a coordinating agent could request information from several specialists. A network agent might examine BGP sessions, routing tables and interface counters, while a virtualisation agent investigates cluster health, virtual machine state and resource utilisation. A monitoring agent could analyse historical telemetry to determine whether the problem represents a new event or part of a longer-term pattern.

The results would then be consolidated into a common assessment, with discrepancies identified for further investigation.

This approach can improve separation of responsibilities and make permissions easier to manage. It can also reduce the amount of irrelevant information presented to individual models, potentially improving efficiency and analytical focus.

Nevertheless, multi-agent architectures introduce their own challenges. Each additional agent creates opportunities for communication failures, inconsistent interpretations and increased execution time. An agent may misunderstand the output of another agent, or multiple agents may independently repeat the same investigation.

For this reason, multi-agent systems should be introduced where there is a clear operational benefit. The objective is not to maximise the number of intelligent components, but to create a system that performs useful work reliably.

In some circumstances, a single model with well-designed tools and a deterministic workflow will be simpler, faster and easier to maintain.

MCP and the Importance of Controlled Integration

The Model Context Protocol has emerged as an important mechanism for connecting AI applications with external systems. By defining a standard approach to exposing tools and information sources, MCP can reduce the need for bespoke integrations and make it easier to connect different AI applications to the same operational environment.

For a cloud infrastructure provider, potential integrations include virtualisation platforms such as Proxmox, monitoring systems such as Zabbix and LibreNMS, customer management applications, vulnerability assessment tools and network configuration repositories.

However, the design of these integrations requires considerable care.

An infrastructure API may expose hundreds of operations, including actions capable of deleting resources, modifying network configurations or changing access permissions. Simply making an entire API available to an AI model would create unnecessary risk.

A more appropriate approach is to expose a deliberately restricted collection of capabilities through an integration service. A virtualisation integration, for example, might initially support retrieving cluster health, listing authorised virtual machines and examining recent task history. Administrative operations could remain unavailable until specific use cases have been evaluated and appropriate safeguards implemented.

This also provides an opportunity to normalise information from different systems. Rather than requiring the model to understand every proprietary API response, the integration layer can present consistent, validated data structures.

MCP itself does not provide a complete solution to authentication, authorisation or information security. Those responsibilities remain with the surrounding application and the systems being accessed.

The protocol should therefore be regarded as an integration mechanism operating within a wider security architecture, not as a substitute for that architecture.

Multi-Tenancy: The Most Important Security Boundary

For organisations providing shared infrastructure services, tenant isolation is among the most important considerations when introducing AI.

A multi-tenant platform may contain information belonging to numerous customers, including infrastructure configurations, performance metrics, service records and potentially sensitive operational data. An AI system must not be able to retrieve information belonging to one customer while responding to a request from another.

This requirement cannot be satisfied simply by instructing the language model to respect customer boundaries.

Language models operate on the information made available to them, and their behaviour can be influenced by unexpected inputs, ambiguous instructions or maliciously constructed content. Security controls that depend entirely on the model making the correct decision are therefore insufficient.

Tenant isolation must be enforced by the underlying application and integration services.

When a customer submits a request, the application should establish the customer's identity through a trusted authentication mechanism. The resulting tenant context must be applied to every subsequent operation, including database queries, document retrieval, monitoring requests and tool execution.

The model should never be allowed to choose its own tenant identifier or override the scope established by the authenticated session.

This principle extends to retrieval-augmented generation, commonly known as RAG. Organisations often use vector databases and document indexes to provide models with relevant internal information. If those indexes contain material belonging to multiple customers, retrieval operations must enforce appropriate access controls before any information reaches the model.

A shared search index without effective tenant filtering creates a potential route for information disclosure. Depending on the sensitivity of the data and the architecture involved, organisations may prefer physically separate indexes, dedicated storage or rigorously enforced logical isolation.

The same concerns apply to cached responses, conversation history, generated reports and diagnostic logs.

A secure multi-tenant AI platform must assume that models can make mistakes and ensure that those mistakes cannot bypass the underlying access-control boundaries.

Reliable Execution Requires More Than Reasoning

An AI agent performing a multi-stage task generally operates through a repeated cycle of gathering information, analysing the results and deciding what to do next.

This process introduces uncertainty because the sequence of operations may not be fully predetermined. Unlike a conventional script that follows a fixed execution path, an agent may select different tools or investigative approaches depending on the information it encounters.

That flexibility is useful, but it also creates new failure modes.

An agent might repeatedly query an unavailable service, misunderstand an error response or continue investigating long after sufficient evidence has been collected. It might also interpret an expected operational event as a serious incident because it lacks relevant historical context.

Loop engineering addresses these problems by placing explicit controls around agent execution.

For example, an agent investigating network congestion should not automatically classify every period of high utilisation as an operational failure. In many environments, scheduled backups or data replication activities generate substantial but entirely expected traffic.

A reliable investigation would examine the duration of the utilisation, compare it with historical patterns and correlate it with packet loss, interface errors or other indicators of genuine service degradation. It should also consider known scheduled workloads before escalating the event.

The agent's conclusions should distinguish clearly between observed facts and analytical interpretations. A monitoring system may confirm that utilisation exceeded a threshold for several minutes, but determining whether that represents a service problem requires additional context.

Execution limits, timeouts, retry policies and explicit completion criteria help ensure that investigations remain bounded. Where evidence is insufficient, the system should report uncertainty rather than inventing a definitive explanation.

These mechanisms are not unique to artificial intelligence. They are applications of established reliability engineering principles to a new category of software component.

Skills as an Operational Knowledge Layer

A significant proportion of infrastructure engineering involves applying established procedures to recurring classes of problems.

Engineers develop troubleshooting methods, operational standards and configuration practices through experience. Much of this knowledge is eventually captured in documentation, although its quality and accessibility can vary considerably.

AI skills provide an opportunity to transform that material into reusable operational capabilities.

A network troubleshooting skill might describe how to examine routing adjacencies, inspect interface status, test connectivity and interpret relevant logs. A vulnerability assessment skill could define how scan results should be classified, correlated with asset information and presented in a standardised report.

The skill does not replace the underlying technical tools. Instead, it provides procedural guidance that helps the AI system use those tools consistently.

This can be particularly valuable for organisations operating infrastructure across multiple locations. A standardised procedure can ensure that similar investigations follow comparable methods regardless of which site or engineer initiates the work.

However, operational knowledge must remain current. A troubleshooting procedure written for an earlier software release may contain commands or assumptions that no longer apply. An AI system following that procedure could produce incorrect results despite executing the instructions exactly as written.

Skills should therefore have defined owners, version histories and review processes. They should be tested against representative environments and updated when underlying systems change.

This approach also supports knowledge preservation. Rather than allowing critical operational experience to remain exclusively with individual engineers, organisations can progressively capture that expertise in structured, maintainable resources.

Human Oversight and Progressive Autonomy

The question of how much authority to grant an AI system should be answered according to operational risk rather than technical capability alone.

Retrieving a monitoring metric is fundamentally different from modifying a production firewall policy. Both operations may be technically possible through the same integration, but their consequences are substantially different.

A sensible adoption strategy begins with read-only access and analytical assistance. Agents can collect information, correlate events, identify anomalies and prepare recommendations without making changes to the underlying environment.

Once the platform has demonstrated reliable behaviour, organisations may introduce narrowly defined low-risk actions. These could include creating draft incident tickets, preparing change requests or updating non-critical records.

More consequential operations should require explicit approval and independent safeguards. Changes affecting production routing, customer services, access permissions or data integrity should be governed by established change-management procedures.

Human approval should also be meaningful. Presenting an operator with an unexplained command and asking for confirmation provides limited protection. The system should explain the proposed action, its expected effect, relevant risks and any available rollback procedure.

Progressive autonomy allows organisations to increase automation while maintaining confidence in operational controls.

The aim is not to eliminate engineers from infrastructure management. It is to reduce unnecessary manual effort and allow technical staff to concentrate on activities requiring judgement, experience and accountability.

Computer Use and Legacy Infrastructure

Not every operational platform provides a suitable API. Some specialist applications and older management systems remain accessible primarily through graphical interfaces.

Computer-use AI offers a potential method of interacting with these systems by interpreting screen content and performing actions through mouse and keyboard controls.

This capability can be useful for repetitive administrative work, information gathering and application testing. It may also provide a temporary integration method while more robust interfaces are developed.

Nevertheless, graphical automation is inherently vulnerable to interface changes and ambiguous visual states. A dialog appearing unexpectedly, a control moving position or a page failing to load correctly can alter the meaning of subsequent actions.

For infrastructure operations, supported APIs and controlled command interfaces will generally provide more predictable behaviour where they are available.

Computer-use systems should therefore operate within constrained environments, with restricted credentials, clear execution boundaries and verification of important outcomes.

The technology expands the range of systems that intelligent automation can potentially reach, but it does not remove the need for conventional software engineering discipline.

Observability: Understanding What the AI Actually Did

Reliable infrastructure depends on observability. Engineers need access to logs, metrics and traces that explain how systems behave, particularly when something goes wrong.

AI platforms require the same capabilities, with additional attention to model decisions and tool execution.

When an agent completes an investigation, operators should be able to reconstruct the sequence of events. This includes identifying which tools were invoked, what information was retrieved, which actions were proposed, whether approvals were obtained and how the final conclusion was reached.

Operational metrics should also capture task completion rates, execution latency, tool failures, retry behaviour and resource consumption. For self-hosted inference environments, GPU utilisation, memory requirements, model throughput and concurrency become relevant capacity-planning considerations.

These measurements help distinguish genuine improvements in operational efficiency from demonstrations that appear impressive but are expensive or unreliable in practice.

Observability must also be designed with information security in mind. Diagnostic records may contain sensitive operational details, customer identifiers or confidential data. Logging systems should therefore apply appropriate access controls, retention policies and data minimisation.

An AI system that cannot provide a meaningful account of its operational activity is difficult to troubleshoot, audit or trust.

AI Governance and Existing Management Systems

As AI becomes integrated into business operations, organisations must consider how its risks and responsibilities fit within existing governance structures.

ISO/IEC 42001 provides a framework for establishing and maintaining an artificial intelligence management system. Its principles can complement established information security and business continuity practices, including those associated with ISO/IEC 27001 and ISO 22301.

For organisations already operating formal management systems, there is considerable value in integrating AI governance into existing processes rather than creating an entirely separate administrative structure.

Risk assessments can address AI-specific threats, including inappropriate tool execution, unreliable outputs, information disclosure and dependence on external model providers. Change-management procedures can govern model upgrades, integration changes and modifications to operational skills.

Incident management processes can be extended to address failures involving AI-assisted operations, while internal audits can examine whether the platform continues to enforce its intended security and operational controls.

Governance should also establish clear accountability. Organisations need to know who owns an AI system, who authorises its capabilities and who is responsible when an automated process produces an incorrect or harmful outcome.

Formal compliance provides a useful management framework, but it does not guarantee technical reliability. Policies and procedures must be supported by effective engineering controls and evidence that those controls operate as intended.

Moving from Demonstration to Production

The difference between an AI prototype and a dependable production platform is substantial.

A demonstration may show an agent successfully investigating a single infrastructure incident. Production deployment requires confidence that the same system can operate across a much wider range of circumstances, including unexpected inputs, unavailable services, incomplete information and attempted security violations.

Testing should therefore examine more than whether an agent can complete its intended task.

An organisation should verify that unauthorised resources remain inaccessible, that tool permissions are enforced correctly and that failures are handled without uncontrolled retries or inappropriate actions.

Adversarial testing is particularly important. Information retrieved from external documents, websites or tool responses may contain instructions intended to manipulate the model. This is commonly known as prompt injection, and it creates risks when an AI system treats untrusted content as authoritative instructions.

A secure architecture must preserve the distinction between data being analysed and instructions governing the system's behaviour.

Operational acceptance criteria should also be measurable. Task completion rates, error frequencies, latency, resource consumption and the need for human intervention provide useful indicators of whether an implementation is ready for wider use.

These measures help organisations make deployment decisions based on evidence rather than enthusiasm.

The Commercial Value of Technical Credibility

The adoption of AI also influences how technology providers communicate their expertise.

As AI-assisted search and answer systems become more widely used, organisations increasingly need to publish information that is useful both to human readers and to automated discovery platforms.

This is where concepts such as Generative Engine Optimisation and Answer Engine Optimisation become relevant.

For infrastructure providers, credible technical content can explain complex services, demonstrate operational experience and help prospective customers evaluate architectural decisions. Detailed discussions of network design, security, virtualisation, resilience and automation often provide greater long-term value than generic promotional material.

The same principle applies to AI itself. Businesses evaluating intelligent automation need clear explanations of what systems can accomplish, how they are secured and where their limitations lie.

Publishing substantive engineering material helps establish credibility because it demonstrates an understanding of the practical challenges involved.

Technical authority is built through evidence, clarity and experience rather than fashionable terminology.

Building for the Future Without Waiting for AGI

Discussions about artificial general intelligence and superintelligence often dominate public debate about the future of AI. Although these topics raise important questions, they should not distract organisations from the opportunities and challenges presented by existing technology.

Current language models are already capable of supporting useful operational workflows when combined with appropriate tools, structured knowledge and reliable engineering controls.

The immediate challenge is not to create systems that possess unrestricted general intelligence. It is to develop applications that perform defined tasks effectively, recognise their limitations and operate safely within established organisational boundaries.

Architectures built around clear interfaces, independent security controls and interchangeable model services are also better positioned to accommodate future developments. As models improve, organisations can introduce new capabilities without abandoning the operational foundations they have already established.

This approach avoids excessive dependence on any single model provider or generation of technology.

It also recognises that the long-term value of enterprise AI will depend on the quality of the systems in which intelligence is embedded.

Conclusion: The Future of Enterprise AI Is an Engineering Challenge

Artificial intelligence is entering a stage in which the capabilities of individual models are only one consideration in a much larger architectural problem.

Harness engineering, agent orchestration, MCP integration, reusable skills and controlled execution loops provide the foundations for systems that can perform meaningful operational work.

Security boundaries, tenant isolation, observability and governance determine whether those systems can be trusted in environments where reliability and accountability are essential.

For cloud infrastructure providers and enterprise technology teams, the opportunity is significant. Intelligent systems can improve access to operational knowledge, accelerate investigations, reduce repetitive tasks and support more consistent decision-making.

Realising those benefits, however, requires the same engineering discipline applied to any other business-critical technology.

The future of enterprise AI will not be determined simply by which organisation deploys the largest or most capable model. It will be shaped by those that successfully combine intelligence with secure architecture, reliable execution and effective operational governance.

Intelligence creates possibilities. Engineering makes those possibilities dependable.

Read more