For a decade, Kubernetes was the right answer. It organized containers, scaled services horizontally and gave platform teams a shared vocabulary for running software in production. It abstracted away enough of the underlying complexity that engineers could stop thinking about servers and start thinking about services. Most cloud-native infrastructure today is built on top of it, directly or in spirit, and EKS made that model the default for the majority of enterprise teams running workloads on AWS.

The workload that defined that era was the stateless HTTP request, fast in, fast out, disposable. A user action triggers a request, the request hits a service, the service returns a response and the container is done. Kubernetes was optimized for that pattern down to the scheduler internals: Bin-pack containers onto nodes, autoscale on CPU and memory, evict and reschedule when something goes wrong. The whole system is tuned around the assumption that individual units of work are short, stateless and interchangeable.

That assumption no longer holds for the workloads that matter most right now.

The agent workload is structurally different

Agents are long-running, stateful processes. They reason across time, call external tools, spawn subprocesses, write and execute code, and make decisions that depend on what happened five steps earlier in the same task. A single-agent workflow might run for minutes or hours, touching a dozen external systems and generating intermediate outputs that subsequent steps depend on. The compute layer for that kind of work needs to do things the old model was never asked to do. That is the new pattern: Execution infrastructure designed around agent semantics rather than request semantics.

The Kubernetes community itself has acknowledged this mismatch. In March 2026, Kubernetes SIG Apps published an introduction to Agent Sandbox, a new CRD-based abstraction designed specifically for singleton, stateful agent workloads. The framing is direct: The ecosystem is moving from short-lived, isolated tasks to deploying multiple, coordinated AI agents that run continuously, and mapping those workloads to traditional Kubernetes primitives requires an entirely new abstraction. The fact that the Kubernetes maintainers built a dedicated primitive for this, rather than recommending teams compose one from existing resources, is itself the clearest signal that agent execution does not fit the old model.

What agent execution actually requires

Concretely, it requires four things. First, isolated execution environments that provision in milliseconds, not minutes, so each agent task gets its own sandbox for code execution and tool calls without blocking the reasoning loop. The difference between a two-second environment and a two-minute environment is not a performance optimization; it determines whether the architecture is viable at all. Second, durable state management across the full task lifecycle, so an agent can pause, hand off or resume without re-initializing from scratch and burning tokens to reconstruct context it already built. Third, coordination primitives for multi-agent work: The ability to spawn subagents, pass structured outputs between them and track task dependencies across a graph of concurrent processes. Production agent systems are rarely single agents; they are pipelines of specialized agents with handoffs that need to be reliable and inspectable. Fourth, credentials and secrets management that travel with the execution context, so agents can authenticate to external services securely without exposing credentials in the task definition, logs or the environment variables of a shared container.

The mismatch shows up fast in production

Kubernetes and EKS expose the mismatch quickly in practice. Pod eviction terminates an agent mid-task with no clean recovery path. Autoscaling reads CPU utilization as the load signal, but an agent holding a long inference connection looks idle to the scheduler even when it is doing the most consequential work in the pipeline. Provisioning a new environment takes 45 seconds to two minutes on a well-tuned cluster; agent workloads need that in under two seconds or the reasoning loop stalls and the user experience degrades visibly. These are not edge cases or misconfigurations. They are the normal operating conditions for production agent workloads running on infrastructure that was not designed for them.

The utilization data makes the broader cost picture even starker. The 2026 State of Kubernetes Optimization Report from CAST AI, drawn from analysis of over 23,000 production clusters across AWS, Azure and GCP, found average CPU utilization at 8 percent, down from 10 percent the year prior. Memory utilization fell from 23 to 20 percent. CPU overprovisioning jumped from 40 to 69 percent year over year. These numbers reflect clusters running traditional workloads, and the pattern is worsening, not improving, as environments scale. Agent workloads compound this problem further. An agent holding an open inference connection or waiting on a tool call registers as idle to a scheduler that reads CPU and memory as the only meaningful load signals. The infrastructure responds to the wrong metric, overprovisioning capacity for demand it cannot measure, while the actual bottleneck, environment provisioning latency and state continuity, goes unaddressed.

Security is not the same problem it was before

Agent workloads change the threat model at the infrastructure level. A compromised stateless service exposes a narrow surface defined by its API contracts. A compromised agent exposes every system it can reach, every credential it holds and every action it is authorized to take on behalf of the user. Agents generate and execute their own code, make non-deterministic tool-call decisions and accumulate context across long-running sessions. Standard container namespacing does not contain that kind of risk. Kernel-level isolation, default-deny network egress, scoped credentials per session and agent-aware observability are not optional hardening steps. They are baseline requirements for running agents in production.

What teams that ship agents have already figured out

Some of the clearest evidence for this shift comes not from infrastructure vendors but from product engineering teams running agents at scale on their own code. In late 2025, Ramp’s engineering team published a detailed account of building Inspect, their internal background coding agent. Each Inspect session runs in a sandboxed VM with a full-stack development environment and deep integrations across their observability, CI, and deployment tooling. The architecture requirements map almost exactly to the four primitives above. Filesystem snapshots keep sessions starting in seconds rather than minutes. Sessions are isolated and stateful. The agent can run tests, review telemetry, query feature flags and visually verify frontend changes in a real browser. And the whole system supports unlimited concurrency, so engineers can spin up ten parallel sessions exploring different approaches to the same problem without contention.

The results speak for themselves. Within months of launch, roughly 30 percent of all pull requests merged to Ramp’s frontend and backend repositories were written by Inspect. That level of adoption was not mandated. It happened because the execution environment was fast enough, capable enough and well-integrated enough that the agent was strictly better than a local workflow for a meaningful share of tasks. The key insight from the Ramp case is not about the model. It is about the execution layer. As their team put it, session speed should only be limited by model-provider time-to-first-token; everything else, like cloning and installing, needs to be done before the session starts. That is a statement about infrastructure, not intelligence.

The ecosystem is catching up, but defaults are sticky

None of that is a criticism of the tools. Kubernetes solved exactly the problem it was designed for, and it solved it well. The issue is that infrastructure defaults are sticky. Teams inherit them, build on top of them and optimize within their constraints long after the underlying workload has changed. The Kubernetes community’s own response, the Agent Sandbox project under SIG Apps, validates the thesis that a new abstraction is necessary. The new primitives the community is building include warm pools for near-zero cold starts, lifecycle management for suspending and resuming idle agents without losing state, and pluggable kernel isolation for secure execution of untrusted code. These are not incremental improvements to existing resources. They are net-new abstractions that acknowledge the old model does not stretch to fit.

But adoption of purpose-built agent infrastructure remains early. Enterprises building agent pipelines today are largely running a request-oriented orchestration model against an execution-oriented workload, and the mismatch shows up in task failure rates, runaway costs and debugging cycles that have no good tooling because the observability layer was also designed for stateless services.

The structural advantage is available now

The infrastructure to close that gap exists now. The prerequisite is recognizing that agent execution is a first-class compute pattern with its own primitives and its own requirements, not a variant of the stateless service model that defined the last decade. Teams that make that shift early will have a meaningful structural advantage. The ones that do not will spend the next two years wondering why their agent systems are unreliable at a scale that should be tractable.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Red Hat OpenShift 4.22, an update to the company’s hybrid cloud application platform, is now generally available. The release focuses on cutting cloud infrastructure costs, simplifying operations of virtualized workloads, and securing sensitive data.

Announced July 14, Red Hat OpenShift 4.22 continues to harden the platform foundation to meet growing security standards, helping reduce the manual effort of compliance and risk mitigation, Red Hat said. The introduction of a minimal Red Hat Universal Base Image (UBI) strips away non-essential packages to reduce the overall attack surface. With Red Hat OpenShift sandboxed containers 1.12, OpenShift 4.22 makes support for confidential containers on bare metal generally available.

The OpenShift 4.22 release also introduces confidential AI as a technology preview. With confidential AI, organizations can isolate and run highly sensitive workloads and proprietary AI algorithms inside a cryptographically isolated slice of memory and CPU, providing data privacy even during runtime execution, according to Red Hat.

OpenShift 4.22 also brings new Red Hat OpenShift Virtualization capabilities. A new Ethernet virtual private network integration with user-defined networks allows teams to connect containerized and virtualized workloads to external infrastructure. Volume groups now can be used to execute multi-volume snapshots for VMs, providing a crash-consistent backup mechanism that simplifies disaster recovery. And the introduction of two-node OpenShift with fencing provides a highly resilient and resource-efficient option for constrained edge environments, Red Hat said.

In addition, OpenShift 4.22 offers new platform capabilities designed to optimize resource usage and lower operational overhead. The Red Hat build of Karpenter, an open source auto-scaler that right-sizes compute instances for Kubernetes clusters, is now generally available for Red Hat OpenShift Service on AWS with hosted control planes. And customers running Red Hat OpenShift Service on AWS with hosted control planes now can integrate AWS EC2 Spot Instances for fault-tolerant workloads to save on costs.

Finally, Red Hat OpenShift 4.22 introduces the JobSet operator to streamline large-scale distributed training runs and LLM fine-tuning. This framework coordinates multiple related jobs as a single unit, maximizing the use of expensive GPU compute.

It’s been several years since generative AI and large language models (LLMs) took the world by storm. LLMs surpassed earlier natural-language systems at generating text, while diffusion models enabled generating images, music, and videos.

These generative AI models work well in the digital world, but on their own, they have limited capabilities to comprehend the three-dimensional physical world and other spaces. This includes the objects occupying an area, how they relate to each other, tracking movement, and answering complex questions requiring an understanding of dimensions, distances, motion, and collisions.

Spatial intelligence is an AI capability that allows models to reason about three-dimensional space. These models can generate 3D scenes of the world and other spaces. This content can then be displayed through traditional renderers, game engines, or AR/VR systems that use spatial computing techniques. But it’s the spatial intelligence model’s ability to connect natural language with 3D models that has the most applications in robotics, manufacturing, construction, and other physical environments.   

Dr. Fei-Fei Li, often called the godmother of AI, published a manifesto on how spatial intelligence is AI’s next frontier, contrasting it with LLMs. “While current state-of-the-art AI can excel at reading, writing, research, and pattern recognition in data, these same models bear fundamental limitations when representing or interacting with the physical world,” wrote Dr. Li. “Our view of the world is holistic—not just what we’re looking at, but how everything relates spatially, what it means, and why it matters. Understanding this through imagination, reasoning, creation, and interaction—not just descriptions—is the power of spatial intelligence.”

The concept of spatial intelligence isn’t new and was described in Howard Gardner’s book, Frames of Mind, in 1983. Recent breakthroughs, including the launch of World Labs’ Marble and its $1 billion funding round, and competing approaches from Google’s Genie 3 and Nvidia Cosmos, should put spatial intelligence and world models on more R&D road maps.

What are spatial intelligence models?

It’s important to develop AI literacy and understand the terminology and concepts related to the physical world and 3D AI technologies: 

• Spatial intelligence encompasses specialized approaches such as geospatial models for mapping the physical world and building information modeling (BIM) for modeling physical structures. It also extends to generative 3D, robotics, and physical reasoning applications.
• World models are a class of neural network architectures and are currently a prominent approach to building spatial intelligence.
• Digital twins are live, virtual replicas of physical assets that combine 3D models with real-time sensor data. Spatial intelligence, an emerging capability of digital twins, adds natural-language prompting, generative scenario exploration, and physics-aware reasoning.
• Spatial computing refers to digital content anchored in and interacting with physical space, sensed and rendered in three dimensions and delivered through AR/VR and mixed-reality systems.

“Spatial intelligence models go beyond pixels to understand the 3D structure of the world—how objects are positioned, how they move, and how they interact,” says David Fattal, founder and CTO at Leia. “This enables applications like more realistic video generation, spatial computing interfaces, and AI systems that can reason about physical environments. As real-world 3D data becomes more available, these models will become foundational to the next generation of visual AI.”

Monitoring the built environment

To better understand spatial intelligence, let’s consider physical infrastructure such as bridges and buildings. The American Society of Civil Engineers estimates a $9.1 trillion investment is needed from 2024 through 2033 to achieve a state of good repair. When maintenance and monitoring lag, it can lead to major failures such as the 2022 collapse of the Fern Hollow Bridge in Pittsburgh.

Spatial intelligence and the development of digital twins may help identify issues earlier and prioritize where investments are needed. “Spatial intelligence models serve as the 4D digital blueprints for our built environment, allowing us to visualize and predict the complex interactions between aging assets and the shifting ground beneath them,” says Patrick Cozzi, chief platform officer at Bentley Systems. “By synthesizing disparate geospatial data into a living digital twin, these models provide the foresight necessary to mitigate the hidden risks of structural fatigue and subsurface instability.”

There’s a significant challenge in bridge health monitoring and transitioning from manual, infrequent structural inspections to leveraging sensors, digital twins, and spatial intelligence. Cozzi adds, “This integration of continuous field data moves beyond static documentation, empowering agencies to evolve from reactive repairs to proactive, resilient asset management that safeguards the long-term integrity of our most critical public systems.”

Avoiding collisions

Bridges are largely static, but the real world is increasingly being occupied by autonomous systems such as self-driving cars, robots, and drones. And where there are moving systems, there is a risk of collisions.

“Spatial intelligence models are AI systems that reason about the physical world by combining vision, sensor data, and contextual cues to understand space, motion, and object relationships,” says Sudeep George, CTO at iMerit. “The value of spatial intelligence models lies not just in perceiving an environment, but in enabling machines to act within it safely and in real time. That is especially important in robotics and autonomous systems, where decisions must be made in complex, multimodal, fast-changing settings.”

To see one example, this tutorial for simulating robotic environments combines Nvidia Isaac Sim, an open source robotics reference framework, with spatial intelligence in Marble from World Labs. 

Today’s collision detection systems, such as those used in autonomous vehicles, typically rely on modules for sensing, perception, planning, and control. Spatial intelligence models may offer improvements by assessing the collision risks of unidentified objects or by tracking objects that move out of sensor view. For example, Waymo’s World Model, built on Genie 3, is a simulator that generates complex weather conditions and other critical safety events.

For an out-of-this-world example, James Urquhart, field CTO and technology evangelist at Kamiwaza, has delivered several examples of spatial intelligence applications, including one for satellite collision detection and conflict analysis. Urquhart says, “Models that specialize in these types of data sets, as well as the physics and geography of the real world, enable faster and more accurate decision-making for tasks that depend on them.”

Applying spatial intelligence

Recent spatial intelligence announcements include creating 3D worlds from image or text prompts with Marble and simulating water physics, lighting, weather, and animal behavior with Genie 3. But difficulties remain in bringing spatial intelligence to physical-world use cases.

“Spatial intelligence and world models are laying the groundwork for future AI agents that will be able to interact in and with our physical world,” says Jason Corso, cofounder and chief scientist at Voxel51. “These models are significantly more challenging to develop and test, largely because the data underlying their development is complex, and it’s hard to handle all of the combinatorics involved in the physical world.”

In addition to learning the models and prototyping with them, development and data leaders need to review the data assets that will feed spatial intelligence models. “Spatial intelligence models translate location signals into a structured understanding of the real world, but they’re only as reliable as the data beneath them,” says Dan Adams, executive vice president and general manager of Enrich at Precisely. “The real unlock isn’t the model—it’s the reference layer with persistent identifiers, confidence metadata, and source lineage that lets AI reason about places, not just match strings.”

Even once applications are developed, there will be infrastructure challenges in deploying them at the edge. Ali Kayyam, principal research scientist at BrainChip, says, “The key to unlocking spatial intelligence at scale is having the low-power, event-driven hardware that can run it at the sensor in real time where it matters most.”

Where to get started

My suggestions for developers looking to get hands-on with spatial intelligence and world models:

• To try out Marble, review their API documentation and case studies, and then experiment with a developer-focused React application.
• Review the Nvidia Cosmos developer hub, documentation, and cookbook of case studies and learning paths.
• You can get an overview of Genie 3, but access is currently restricted through Project Genie, which requires a Google AI Ultra subscription.

The model of the forward deployed engineer is sweeping through enterprise IT like a gold rush, and I’m concerned that many companies don’t understand what they’re signing up for.

Let’s start with the headline numbers. AWS announced a $1 billion investment in a new Forward Deployed Engineering organization. Google Cloud committed $750 million to expand similar programs. Microsoft has been running Azure-focused embedded engineering teams for years, including partnerships with Accenture to scale forward deployed engineering practices. All three are pitching the same story: We’ll send engineers to work directly with your teams, help you deploy AI, and accelerate your digital transformation. You get top-tier technical talent for free, and we get to partner with you on your journey.

It sounds reasonable on the surface. It sounds collaborative, even generous. But I’ve been in this industry long enough to know that when a multi-billion-dollar company offers you something for free, they’re sure to get much more than they give.

What you actually get

The forward deployed engineer model isn’t new. The consulting industry has been doing some version of it for decades. What makes this different is the scale and the direct financial incentive behind it.

These engineers work for the cloud provider. They’re not your employees. They’re not independent consultants. They’re technically excellent professionals who are being paid to solve your immediate problems while simultaneously building relationships and architectures that favor their employer’s ecosystem. Think about it from their perspective. Those forward engineers are evaluated on whether customers succeed with their employer’s platform. They’re rewarded when enterprises adopt more services from that platform. Their career advancement depends on making AWS, Google Cloud, or Microsoft Azure the obvious choice for all of your technical decisions.

This isn’t a criticism of the individual engineers. Many of them are genuinely talented and genuinely want to help. But they’re operating within a system that rewards specific outcomes, and those outcomes align with the vendor’s financial interests, not necessarily yours.

The problem no one talks about

Here’s what I see happening at enterprises right now. A company decides they need help deploying AI. A cloud provider offers to embed engineers at no additional cost. Those engineers work alongside internal teams, make architectural recommendations, and help build out systems. Six months later, the company has a production AI system running on a single cloud platform, built by people with deep expertise in that specific platform.

The problem? Nobody evaluated whether that platform was actually the best choice for the business. Nobody looked at alternatives. Nobody asked whether a multicloud architecture or best-of-breed approach might deliver better results at lower cost.

The engineers embedded in these programs are not going to recommend that you split your workloads across providers. They’re not going to suggest you use open source tools where they make sense. They’re not going to point you toward a competitor when their employer’s solution will work well enough. That’s not how these programs are designed to function. What you’re getting is optimized architecture for a single cloud brand, not optimized architecture for your business.

The financial reality will hit

The bills are going to come due, and they’re going to be painful. I’ve watched this pattern play out before. When enterprises lock into a single cloud provider through these embedded engineering programs, they often discover two or three years later that they’re paying premiums that their more independent-thinking competitors avoided.

The reasons are straightforward. When you’re architecting systems around a single platform, you naturally fall into usage patterns that favor that platform’s pricing structures. You use their managed databases instead of portable alternatives. You adopt their AI services instead of evaluating third-party options. You build workflows that only work within their ecosystem. And when it comes time to renegotiate or benchmark against alternatives, you find that migrating would cost more than accepting whatever pricing they offer.

I’ve spent the past decade helping companies untangle from these situations. I’ve seen organizations with cloud bills 15 to 20 times higher than they should be, unable to migrate because their entire AI infrastructure is built on proprietary services that only work on one platform. The forward deployed engineer programs are accelerating this problem. They’re making it easier to get into these situations and harder to get out.

Think before you commit

Before you accept one of these programs, consider these three recommendations.

First, require independent architecture oversight from day one. Hire or engage architects who work for your company, not for your cloud provider. They should evaluate every recommendation made by embedded engineers against business requirements and compare options across providers. This isn’t about being suspicious of the engineers. It’s about ensuring that decisions are made with your interests in mind.

Second, demand a clear exit strategy before you begin. Ask the cloud provider to document which proprietary services you’re using, what migration paths exist, and what the cost would be to move to an alternative platform. If they can’t provide that information, or if the migration costs seem impossibly high, that’s a sign that you’re building technical debt that will be very expensive to service later.

Third, benchmark your costs continuously. Set up internal processes to compare your cloud spending against industry benchmarks and against what your competitors might be paying for similar workloads. Don’t wait until your contract renewal to discover that you’re paying premium prices. Monitor expenses from the beginning, and be willing to challenge your cloud provider if you’re not getting value that justifies the cost.

The bottom line

The forward deployed engineers are solving real problems. Enterprises genuinely struggle with AI deployment, and having experienced engineers available to help is valuable. I’m not suggesting these programs are fundamentally bad. However, they’re being marketed as neutral partnerships when they’re actually strategic sales programs designed to lock enterprises into specific platforms. The helpful engineers showing up at your office are building dependencies that will be very difficult to break. The “free” technical assistance is being funded by margins on services you’ll be buying for years.

Go in with your eyes open. Use these programs but add your own independent oversight. Build architectures that you could leave if you needed to. And don’t let the immediate satisfaction of having problems solved today blind you to the financial consequences that will arrive tomorrow.