AI agents operate at speeds that far exceed a human's ability to track every action they take. As their role within organizations expands, the problem is no longer enabling them to perform more tasks, but rather knowing when to allow them to continue, and when human intervention becomes necessary.
For James Hodge, chief data strategist and field technical director at Splunk, a Cisco subsidiary, the answer starts with risk management. An organization handling passport data, for example, may choose to shut down the system at the slightest sign of risk, while a low-sensitivity service can afford more leeway before intervening. The difference is not determined by technology alone, but rather by the nature of the data, the sector, the regulatory framework, and the acceptable level of risk.
Expected versus deviant behavior
One of the most difficult challenges that arise with AI agents is that error does not always take one form. It may be the result of a flaw in the system, instructions that were not formulated accurately, external manipulation, or an actual security breach. Hodge said during an exclusive interview with Asharq Al-Awsat on the sidelines of the Splunk conference held in Denver, USA, that the first step should not be to jump directly to the question of whether the incident was security or operational, but rather to find out whether the agent is still acting within the limits expected of him.
He explains that this requires three interconnected layers, starting with a data layer that the agent relies on, a context layer that helps understand what is happening, and then a policy layer that defines what is considered acceptable behavior. Then it is possible to compare what the agent actually does with what was expected of him, and any deviation can be detected before it turns into a bigger problem.
Deviation does not necessarily mean an attack. The agent might do what was asked verbatim, but in an inappropriate manner, because the instructions themselves were not precise. The error may be the result of incomplete programming, or an external attempt to tamper with the agent. Here begins the investigation process to determine whether the problem should be referred to the systems operation team, the security operations center, or in the future to other agents specialized in investigation and analysis.
There is no independence without explanation
Hodge sets a basic condition before talking about giving AI agents greater powers: that the organization can explain what they are doing. He points out that an organization that cannot explain what an agent is doing, what systems it is interacting with, and why it took a certain path, will find it difficult to determine the level of risk, or decide when a decision needs human intervention.
He gives the example of programming tools supported by artificial intelligence, explaining that if a developer uses tools such as “Claude” or “Codex” to create software code, it is important that he actually understands what was written inside it, rather than just accepting the result.
The same principle applies to artificial intelligence agents. If the organization is unable to explain what the agent is doing, it becomes difficult for it to determine whether it can be allowed to implement directly, or merely provide a recommendation that requires human approval. Thus, interpretability becomes part of risk management itself, not just a technical feature. The higher the degree of visibility into an agent's behavior, the easier it is to determine how much autonomy to give him.
Reconstruct the decision from scratch
The problem becomes even more difficult when the agent does not act alone, but rather calls on other agents or systems in a long series of actions. According to Hodge, agent monitoring tools allow recording the claims that enter the system, the inputs they received, the steps they executed, and the outputs that resulted from them, allowing the path that led to the final decision to be reconstructed.
To explain this more clearly, he says, for example, if an employee requests to obtain employee salary data, in a controlled institutional environment, a protection layer can reject the request before it reaches the agent if it violates the policy, and then the monitoring tools continue to record what happens later. This is important when investigating errors, because the organization needs not only to know that an inappropriate decision occurred, but to determine its original source, and whether it resulted from user instructions, an intermediate step, or a decision made by another agent within the workflow.
Visibility from chip to business level
In complex environments an investigation may need to track what is happening across multiple layers, from the chip, network, and infrastructure, to applications, services, and then the business level.
Hodge points out that Splunk's strategy depends on collecting data from various parts of the organization and linking it, citing "Cisco Data Fabric" and "Machine Data Lake" and the capabilities of unified search across data. The goal, he believes, is for an organization to be able to see the original request, then track its impact across the network, infrastructure, application, service, and business process, rather than dealing with each layer in isolation.
This type of insight becomes more important the more agents are able to carry out a series of actions independently, because an error may start in one layer and then have its impact elsewhere entirely.
An internal agent is not the same as an external agent
Hodge highlights a fundamental difference between agents that an organization operates within its technical environment, and external agents that connect to its services via APIs, or other communication protocols.
If the agent is within the boundaries of the organization, it is possible to know what device it is running on, who created it, what tools it uses, and collect detailed data about its behavior. However, if the agent works outside the environment controlled by the organization, visibility may only be available to the request that reached the system, without knowing what happened inside the agent before that. This difference places practical limits on the idea of comprehensive surveillance, especially with the expanding use of third-party agents.
Speed dictates lighter tools
An agent may perform many more calls and actions in seconds than a human user, making it expensive and slow to monitor each step using a large language model. This is why James Hodge says that Splunk relies on a combination of methods, including smaller, task-specific linguistic models, along with statistical techniques and traditional rules, to detect deviations quickly and at a lower cost.
In one of the presentations at the conference, Hodge pointed to an example in which the risk score exceeded 70 percent, which gave strong confidence in the existence of a security incident, but the final decision on whether to stop or allow the work to continue remained with the analyst. He emphasizes that this percentage is not a fixed rule. An organization may choose to treat a much lower risk level as sufficient cause for suspension if an agent is dealing with a critical system, while a greater margin may be accepted in a low-sensitivity service.
Independence is not one level
Hence the concept of “adjustable autonomy”. When a new agent is discovered, an organization begins to understand what it does, what systems it interacts with, what data it requests, and where that data goes.
A risk profile can then be built for the agent, and a policy can be established that specifies whether it is allowed to recommend only, execute with human approval, or operate more autonomously. Hodge believes that this flexibility is necessary, because risk levels vary radically from one use to another, and a single model cannot be applied to all artificial intelligence agents.
Man is still necessary
Despite the trend toward greater independence, Hodge does not consider that institutions have reached a stage where humans can be removed from sensitive decisions. He said human intervention “should stay in place for now,” especially in the most sensitive sectors, because organizations are still learning how to measure the risks, governance, and costs associated with artificial intelligence. He pointed out that the challenge is not limited to building agents capable of implementation, but rather the ability to give the organization's teams sufficient confidence to understand what happens when these systems operate at machine speed. Thus, the transition to broader independence will not happen all at once, but rather through stages that begin with monitoring and understanding, then granting limited powers, reaching higher levels of implementation when the institution becomes able to assess risks more accurately.
Saudi Arabia... regulation helps accelerate
On the Saudi side, Hodge describes the Kingdom as “one of our fastest growing markets,” pointing to the rapid development of the technology sector and the expansion of artificial intelligence projects there. He notes that the presence of clear regulatory frameworks around security, data sovereignty, and the deployment of artificial intelligence technologies represents a factor that helps organizations move more quickly. According to Hodge, the clarity of the rules allows companies, banks, and regulated institutions to know the limits within which they must operate, instead of entering into projects without a clear vision of the acceptable level of risk, or the type of data that can be used. It is noteworthy that this clarity allows Splunk to work with financial institutions in the Kingdom on the basis of existing regulatory frameworks, rather than building new rules for each case. This does not mean that Saudi Arabia gives agents greater independence than other markets. Rather, Hodge says that human intervention is still present in most sectors, and that the decision is related to risks and the regulatory framework more than to the market itself.
Sovereignty brings AI back to data centers
The debate in the Kingdom extends to where to operate these systems. Sending all data to external cloud environments may not be a suitable option in sectors such as banks, national infrastructure, and organizations that handle sensitive data. This is one reason why providing AI and monitoring capabilities within data centers themselves is important for regulated industries, Hodge said.
He pointed to the cooperation with NVIDIA, and the local operation solutions that allow the deployment of agent monitoring tools and artificial intelligence assistance within the organization’s environment, instead of relying exclusively on the public cloud.
But he clarified that this demand is not limited to Saudi Arabia, as there are similar trends in Europe regarding data sovereignty and the place where artificial intelligence agents operate, while he believes that the Kingdom is characterized by rapid development and clarity of policies.
The biggest challenge facing Saudi institutions
When Asharq Al-Awsat was asked about the biggest challenge facing Saudi institutions as they move from artificial intelligence experiments to actual use, Hodge did not put the power of models or the availability of computing at the top of the list.
Rather, he said, the main difficulty is the ability to see data across the entire organization. As he describes it, an organization needs to know what's going on in every system, every application, and every agent, and then present that information in a way that helps the employee assess the level of risk.
This shows that this represents a change in the skills of the IT teams themselves. In traditional systems, it was often possible to determine the status binary: the system is working or not, and the signal is green or red.
In the world of artificial intelligence agents, there is a broader area of uncertainty. The system may appear sound, but it is gradually moving toward a higher level of risk, requiring the employee to interpret trend and context rather than rely on a fixed operating condition.
That's why Hodge cautions against jumping straight into the most advanced AI applications before building the foundation for them. The data layer, context, policies, and cross-platform visibility are not components that come after agent deployment, but rather conditions that must precede the transition to operation at scale.
With the start of the back-to-school and college season, finding the right laptop becomes a top priority for students and parents. These computers are no longer just tools for browsing the Internet or writing traditional text papers, but rather have turned into scientific centers that meet the requirements of students in various disciplines, especially Science Technology Engineering Math STEM projects that require superior computing capabilities and intelligent processing of data and graphics. At the same time, there is an urgent need for an enjoyable period of relaxation and entertainment after the end of long school hours, which requires laptops supported by integrated graphics processors that combine academic efficiency with absolute speed in games.
Choosing a laptop relies on advanced graphics processors and gives students a true shortcut to a device specifically designed to handle challenging coursework, creative projects, and artificial intelligence applications. Thanks to the power of these integrated processors, students can complete their academic tasks much faster than on traditional devices, whether it is searching through notes, training artificial intelligence models, using artificial intelligence agents, generating images, editing video presentations, or creating engineering and programming presentations.
In this topic, we mention a set of tips for choosing advanced laptops that help in studying more effectively, and some new devices in the Arab region that target students.
Artificial intelligence techniques for study and productivity
The demands of modern education require advanced laptops capable of supporting a wide range of activities, from basic productivity applications to scientific software, data analysis and advanced graphics. It is recommended to acquire devices that support AI-accelerated applications, which allows students to complete their tasks more quickly and efficiently and easily handle intensive work that may slow down regular computers. This is done by running large language models (LLMs) and small language models (SLMs) locally on the same device. This allows achieving superior performance levels of up to 30 times in developing artificial intelligence applications and 6 times in artificial intelligence image generation, in addition to providing 10 times faster performance when running demanding games compared to traditional devices.
Below we review the most important advanced and useful features for students that rely on artificial intelligence technologies in the NVIDIA GeForce RTX 50 Series graphics processing units integrated into modern laptops:
• NVIDIA Broadcast works to develop the online learning experience by improving the quality of audio and video when attending lessons remotely or working with other students, which makes group projects and virtual lectures more interactive.
• Players can develop their gaming experience using these computers that use artificial intelligence techniques to increase the frames-per-second FPS thanks to the Deep Learning Super Sampling DLSS technology, which works to generate more images per second in advanced games, providing faster performance than other laptops in the same price category, while maintaining graphics clarity at the highest settings.
• The NVIDIA Jetson Platform is designed in the field of embedded computing for artificial intelligence to provide the higher education and scientific research sector with high-performance, low-power computing capabilities for deep learning and computer vision. This platform allows students and researchers to run fully generative artificial intelligence models on their personal devices and benefit from advanced tools and software to develop robots and intelligent systems capable of independent thinking and learning.
• Reflex technology contributes to reducing the response time for commands from the moment the button is pressed to the moment the result is displayed on the screen in less than ten milliseconds, thus maintaining the player’s superiority and progress in the latest competitive games.
• The NVIDIA NIM feature allows students to create graphics based on artificial intelligence efficiently and at high speed.
• The RTX Chat feature provides a built-in smart assistant that summarizes notes, organizes class schedules, and brainstorms to generate ideas.
The student needs computers that support many important programs for the educational process, which include Ansys Discovery, 3AMR, V, Blender, Adobe, CuPy, CorelDraw Graphics Suite, Unity, MicroStation j, Notch, PyTorch, CapCut, SketchUp, TensorFlow, Topaz Labs, Windows ML, Unreal Engine, RAPIDS, Red Digital Cinema RedCine-X Pro, and Rhino 7. SolidWorks, Streamlabs, VTube Studio, R, Wondershare Filmora, AI Content Ninja, Numba, OBS Studio, Enscape, and other virtual and augmented reality, AI, programming, and editing software for HD video and images.
Designs that combine performance and portability
Students also need laptops that are lightweight and easy to move between lecture halls, home, school, university, or any other place of study. The laptops mentioned in this topic have designs of low thickness and weight, and batteries provide a longer life of up to 40 percent compared to previous generations, which enhances student productivity throughout the day without interruption for recharging.
Despite the superior performance offered by these computers, they remain quiet and cool even under intense pressure thanks to Max-Q technologies powered by artificial intelligence, which improve thermal performance and ensure high operating efficiency across various usage environments.
Absolute speed and creativity
The Gigabyte Gaming A16 computer offers an impressive combination of performance and practical design suitable for students, especially engineering students and creative people:
- Screen: The device features a 16-inch screen with a resolution of 1200 x 1920 pixels using IPS technology, a brightness of 300 nits, and a panoramic display with ultra-thin edges, with a fast refresh rate of 165 Hz to ensure extremely smooth display of visual content and vision.
