AI leads to a platform engineering revival at KubeCon NA 2025 – News

As you would expect this year, some of the conversation at this week KubeCon/CloudNativeCon North America 2025 in Atlanta felt a little bit like a support group. We’re all trying to get past the hype, and come to grips with the risks and opportunities artificial intelligence presents for the cloud native development community.

However, there’s also a positive undercurrent of platform engineering revival going on here, since we can assume most developers and operators are already using some AI-enabled tooling in order to deliver some form of AI-powered application functions.

What are the new ideal practices for equipping cloud native developers with everything they need to build with AI on scalable and resilient infrastructures, including Kubernetes and all of its surrounding ecosystem of both mature and emerging projects and vendors?

“Cloud native and AI-native development are merging, and it’s really an incredible place we’re in right now,” Chris Aniszczyk, chief technology officer of the hosting Cloud Native Computing Foundation, said in the opening keynote. “How do we take all the capabilities Kubernetes provides like autoscaling, and apply that to AI training, and inference, and agents to take it even further?”

Of particular note at this event was the launch of CNCF’s Certified Kubernetes AI Conformance Program, which sets open standards for making AI workloads predictably deployable on Kubernetes so they can be interoperable and portable across different infrastructure types. 

To get our arms around consumption and cost concerns, new Dynamic Resource Allocation capabilities optimize the performance of AI workload deployments across graphics processing units, tensor processing units and other hardware, including mainframes. Yes, AI is finally becoming a responsible cloud-native citizen.

Platforms will change, but the practice of platform engineering grows

At my first KubeCon in 2018 in Seattle, I marveled at the dozens of emerging tools forming around the Kubernetes project. By now, all of those projects should have graduated, or perhaps gotten archived. Such simpler times compared to today’s dizzying CNCF Landscape eye chart of hundreds of projects. 

January 2025’s “DeepSeek Moment” was really just a blip in computing history, but when Chinese researchers produced a solid open-source large langauge model, it became clear that despite infinite capital investment in commercial AI solutions, there will never be one perfect model to rule them all. The open source community can and will build comparable platforms for generative AI and agentic AI development.

“I think it’s a sign of maturity that even when you mix in cloud native development with the current AI hype, organizations are actually realizing that developers are still the internal customers who need platforms as products. You can’t just say ‘get on with it’ and hand them a box of tools,” said Daniel Bryant, head of marketing, Syntasso Ltd., who was distributing a handy little O’Reilly guide to platform productization at the show.

Pavlo Baron, co-founder and chief executive of Platform Engineering Labs, said that on day one, when developers are writing code and making changes, they’re at an abstract level. “They need a concrete representation of the infrastructure, but they don’t want to know anything about networks and databases,” he said. “But on day two, that’s where life begins. Sometimes ops just needs to go into the AWS console, make a quick change, and doesn’t need to care about all the processes. The thing that matters is the focus. You need to see only as much as you need in the platform, because there’s so much noise, and so many signals. “

At the event they introduced their Formæ open-source IaC platform, which can discover and unify all infrastructure-as-code changes in an extended environment as abstracted, stateful elements, so developers can declaratively build atop infrastructure, and operators can make patches and updates without touching Terraform code.

As it turns out, participation in open-source innovation is a highly sought after job experience when hiring platform engineers, even if the direct value impact of such contributions weren’t readily apparent on the employer’s balance sheet.

“We constantly evaluate what’s in the open source catalog for our own use, and we contribute back to open source for the things we modified when it made sense,” a senior platform engineer of a major social network said in a roundtable. “We’re finding that when we hire people who contribute to Kubernetes and related projects, we can bring them in and have them be productive on day one versus a three-to-six month ramp up time.”

Optimization at the network and infrastructure layer

“We’re 10 years into Kubernetes and the foundations are starting to break, and we have to really rethink every layer of the system down to the lowest level,” said Alex Zenla, founder and CTO of Edera Inc., provider of a hardened runtime for lean container environments. “People are getting a little tired of all the tools and all the various things that they have to install to make Kubernetes usable, when it should work out of the box and actually function.”

For instance, Zenla mentioned that Kubernetes doesn’t come with a default networking solution. Different teams can decide to use Celium or Calico, but when they get a distribution from a leading cloud provider, it may contain custom container definitions, networking, permissions, observability and pod scaling implementations, some of which may rely on a specific vendor.

The original CNI network spec is also 10 years old now, and the community keeps finding new ways to simplify network headaches. Tigera Inc., for instance, introduced a new Calico AI assistant to provide natural language query capabilities for troubleshooting network and policy issues in large clusters, as well as bundling observability tools and a hardened operator for ingress and an Istio ambient service mesh within its packaged offering.

Floxdev Inc. demonstrated its unique “uncontained” Flox environment images based on Nix that can take in CI builds and statefully run inside Kubernetes on bare metal, with zero-byte container overhead for instant starts and fewer vulnerability handholds.

Golem (Ziverge Inc.) was there with an agentic runtime orchestration platform for agents that can audit operations and maintain failover state for tens of thousands of discretely identified AI agents inside a single node (through WebAssembly, or WASM, actually). So, if a server crashes or an API response times out, agents can pick up where they left off.

Making the AI-assisted development landscape safer

According to the latest Cloud Native survey, software supply chain attacks are on the rise, and the number of reported CVEs has increased year-over-year by more than 16%. Fortunately, the OpenSSF working group has gathered a technology audit fund and contributors from leading companies that are auditing and finding issues, especially in the the many noncore components of Kubernetes that go into live applications.

With the amount of code moving through the system with AI coding assistance, speed is killing releases without appropriate safety guardrails. We’re even seeing new forms of malicious “slopsquatting” of AI-hallucinated libraries and vibecoded vulnerabilities in packages that the community must address.

“It’s really hard to control the ingredients that are going into our software,” said Glenn Weinstein, CEO of Cloudsmith Ltd. “Over 90% of every build is someone else’s software — open-source libraries, Docker containers from public repos, LLM datasets and OS dependencies. It’s a complete Wild West of dependencies, and DockerHub, PyPi, npmJS, are loaded with bad actors who are slipping in vulnerabilities.”

The vendor now offers an AI-driven software-as-a-service artifact curation and management service, such as a content distribution network that distributes cloud-native packages to developers and an MCP server for agents, scanning registries for vulnerabilities or license issues, then proxying, caching and serving up source package registries before they are incorporated into software.

Kusari Inc. was there offering a SaaS-based platform based on its widely adopted guac open source security tool for inserting dependency and vulnerability checks within the git workflow of every developer pull request, including a new autonomous AI code reviewer that auto-scans code for bugs, security risks and even license risks to ward off ever-present patent trolls.

AuthZed Inc. builds a graph of relationships between people and people, people and data, and data and data within its open-source SpiceDB, so enterprises can make more informed authorization decisions across a variety of systems without having to write new computer programs for policies. And of course, there’s an MCP server so coding agents can consume the service.

Improving observability workflow with AI SREs along for the ride

As much as we try to avoid Gartner buckets, it has coined a new product category called “AI SRE,” for site reliability engineering. It includes AI-assisted systems that can detect, investigate and remediate production issues such as faulty configurations and out-of-control cloud costs.

ControlTheory Inc. recently put out its open-source Gonzo terminal user interface that can monitor and buzz through real-time logs from Kubernetes clusters and other sources to help troubleshoot live issues directly from the command line. Of course, if you want further filtering and tagging of issues, its AI-powered Dstl8 product can display trend analytics and assist SRE investigations. 

Komodor Inc.’s just-announced “Klaudia” agent provides an interactive investigation chat interface and an automated remediation process which can even do “autonomous self-healing” of errors, so human SREs can choose to auto-resolve certain frequently encountered issues and out-of-bound policy mistakes within the environment, while optimizing pod and node usage for cost control and utilization targets.

Getting to the bottom of critical runtime issues with AI will still require teams to shift agentic testing left in new ways — a subject I recently covered.

“People are finding out how easy it is to generate code, and use AI to generate tests, and they forget that code still has errors, just like human code,” said Ole Lensmar, founder and CEO of Testkube, a Kubernetes-native continuous testing platform that recently added an MCP server and the ability to kick off multi-agent, multi-region distributed test suites. “Are the tests relevant? Are they toxic or non toxic? Are they on topic, and how are the test results skewed if the underlying model changes?”

The Intellyx take

Surprisingly, the weather in Atlanta started out much frostier than previous Kubecons in unseasonably warm Novembers in venues such as Detroit and Chicago. However, the warm collaborative atmosphere of the cloud-native community quickly melted any concerns about the future of the Kubernetes ecosystem.

If there’s one thing I can take away from this conference, it’s that this now 10-year-old cloud native community is going to be just fine, and continue to innovate new ways to stay ahead of the pressures of AI development, and whatever comes next. 

All of the goodness of microservices and automation and application programming interfaces that preceded Kubernetes, and an ever-widening river of enabling projects and technologies since, will continue to provide immense scalability and resiliency to the applications of the future — whether or not AI is on the label, or under the surface.

Jason English is a principal analyst and chief marketing officer at Intellyx. He wrote this article for News. At the time of writing, none of the companies mentioned here are Intellyx customers. No AI was used to generate this content. CNCF covered the analyst’s attendance cost for the event, a standard industry practice. ©2025 Intellyx B.V.

Photo: CNCF