The AI ecosystem is still in its infancy. This is evident from the regular releases of immature, yet highly imaginative, open-source solutions. It’s up to companies like Red Hat to anticipate these developments and help such tools mature into enterprise-ready products. “At Red Hat, we take a project, stabilize it, and turn it into a product. We’re going to do the same for AI.”
These are the words of Natale Vinto, Technical Director of Hybrid Platforms at Red Hat. An enterprise IT company’s starting position vis-à-vis open-source maintainers is multifaceted. Of course, companies themselves contribute to ongoing innovation (think of IBM and Red Hat’s Lightwell or Anthropic’s open-sourcing of MCP), but their role is often a moderating one.
From OpenClaw to OpenShift
A prime example of this is OpenClaw. It was released in late January of this year and was, to put it mildly, incredibly dangerous. The highly potent and deceptively simple set of “claws” meant that LLMs, previously typically confined to predefined roles, could suddenly perform all sorts of tasks. Claws are risky, and OpenClaw was exactly that. AI went from being a “brain in a jar” to a brain that could suddenly act on its own, says Burr Sutter, Director of Developer Experience at Red Hat.
Anyone in a business environment handling all kinds of sensitive data who wanted to take advantage of this high degree of agency had to wait a bit. But not as long as you might think. Less than two months later, Red Hat followed up with “Bring Your Own Agent” (BYOA) for Red Hat AI. The idea is to deploy an agent within your existing, trusted IT infrastructure with familiar governance and isolation.
Vinto explains how Red Hat can pivot quickly to achieve something like stabilizing OpenClaw for enterprise use. “I know the pace [of AI innovation] is lightning-fast, but we’re fortunate that we’re not starting from scratch.” According to Vinto, decades of experience in hardening operating systems, platform engineering, application development, DevOps, and more provide the expertise needed to ground new developments in a solid paradigm.
It helps that Red Hat is involved in the open-source community. Vinto describes the company as “the de facto leader in the open-source AI community” due to contributions to vLLM, llm-d, and Kubernetes, among others. “For AI, we understand the need to sandbox agents and to manage security, identity, and authorization for these agents. We therefore incorporate the agent lifecycle into our products, connected to everything needed to run an agent in production.”
AI is (not) new
The idea behind BYOA, the mature version of what OpenClaw promised, is to achieve reproducible results in production. That is a rare thing in the AI domain, where (partly due to a lack of in-depth AI knowledge within organizations) the technology’s promises often outpace its actual practical utility. The technology itself can achieve a great deal, but it needs a clear direction that is executed effectively.
Experimentation is certainly part of that, though. “You can experiment with OpenClaw, but we sandbox it,” says Vinto. “First of all, we limit its network traffic. We place it in a namespace. It’s only your user.” Vinto explains that if an attacker were to take over this OpenClaw instance for any reason, they would only have gained access to the container. For developers, this means a degree of freedom, but with guardrails.
One of the issues surrounding AI revolves around cost. Research on this topic is scarce, but we get the impression that it’s human nature to go for the best available option in a drop-down menu. In the context of AI, this means that users with unlimited token budgets can systematically opt for massive, but extremely expensive, LLMs even for the simplest tasks. Vinto explains that semantic routing counteracts this on the Red Hat platform. The vLLM Semantic Router interprets prompts before they are sent to an LLM and selects the appropriate model based on complexity. “Tokenomics” are thus automated, ensuring that a simple prompt, one that can be answered directly with a single knowledge base quote, does not, so to speak, invoke Claude Fable 5.
Many layers
More than ever with new innovations, AI is impacting all layers of the IT infrastructure. Take Kubernetes as a counterexample. It has brought portability and abstraction where previously painful configuration and incompatibility dominated. But it is fundamentally a victory for IT professionals who operate a layer deeper than most of their colleagues in the field. Meanwhile, AI is relevant not only to that infrastructure layer but also to the user experience, resource optimization, and integrations between applications. Where deterministic scripts once formed a network path between solutions, integrations are now more complex than ever. Agents query each other, issue commands to one another, and orchestrate interactions above the SaaS layer. All of this sounds great, but the IT foundation must be solid to realize these scenarios without all sorts of unintended consequences.
Red Hat’s vision is that any model, any agent, can run on any hardware or cloud. The management layer must remain centralized. Burr Sutter explains that AI should be embedded within a kind of “OS,” just as other user-facing IT already runs on an operating system (often Red Hat’s).
Specifically, Sutter characterizes the “harness,” the framework into which an LLM is placed and can be useful, as the “AI OS.” Strikingly, we’ve seen many AI-related systems described as an AI OS, from data platforms to identity managers. Rarely has the comparison been more apt. An agent harness is viewed as the AI OS by more parties than just Red Hat, Sutter explains. You can draw a comparison between various components. Both an OS and a harness manage a central processor (the LLM in a harness, the CPU in an OS), memory (both short-term and long-term), and device drivers, which serve as tools for an AI model via MCP.
As an expert in Linux, Kubernetes, and enterprise IT architectures, Red Hat feels compelled to respond to this concept of agents with harnesses as manageable, isolated systems through offerings such as Red Hat AI, OpenShift and Red Hat Ansible Automation Platform.
Continuing to make choices
Anyone who closely follows AI can sometimes feel lost or even dizzy. It seems like there’s a new LLM every week, and every month brings a new standard, protocol, or hype. “Prompt engineering,” for example, already feels outdated, while we’ve also recently seen trends like “tokenmaxxing” and “agent loops.” This raises the question of how Red Hat not only responds to something like OpenClaw but also takes into account the fickle nature of AI sentiment in the process.
The answer lies in Red Hat’s portfolio. For instance, Red Hat introduced AI 3.4 Model-as-a-Service as a governed interface for running curated LLMs, allowing admins to track usage and enforce policies. Red Hat AI contains a bootable image with AI libraries and hardware-optimized inference for Nvidia, Intel, and AMD, in addition to enterprise support and IP protection via Open Source Assurance. It offers certainty where there may still be many AI-driven uncertainties.
When asked, Vinto acknowledges the gap between AI hype and adoption. He notes that this gap is also regional. Not entirely surprisingly, users in California are much more eager when it comes to the latest AI solutions, and it comes as no surprise when we hear once again that Europe is somewhat more cautious. This is partly due to concerns about issues such as AI-driven job losses, something Vinto sees as an exaggerated fear. In addition to existing jobs that may be disrupted, he also sees room for entirely new full-time equivalents (FTEs) in a world with AI.
Conclusion: plenty of opportunities, including for open source
We’ve already seen that, for example, established SaaS players and IT hardware vendors react very differently to the AI hype and are influenced by it in different ways. What we find most interesting are the vendors who certainly had a significant say in the “classic” IT era but now have to redefine their role. The mindset is also shifting in relation to open source. For instance, we shouldn’t overlook the fact that an AI building block like vLLM or llm-d is truly open source, whereas “open-source” AI models are merely downloadable files that you can run anywhere, rarely accompanied by information about the data the AI lab used to train them.
“Open source isn’t going away,” Vinto believes. “It’s being transformed. The projects will be different, but open source will remain. The business case for freedom of choice in the tools you need, and for interoperability and open standards, that will always be there.”