Platform engineering 2026: what changed in DevOps, what to skip

Platform engineering 2026 is less a new tool than a new default: most teams now give developers a shared, supported way to build and ship, and then add AI on top of it. The changes that matter this year are a retired ingress controller, a billing reversal on CI, AI agents that join incident work, and a steady shift of effort from writing pipelines to running a platform. Much of the rest is noise. This note sorts the two, for engineers deciding what to practise.

The change that breaks things: Ingress NGINX is gone

The community Ingress NGINX controller was retired in March 2026. There are no more releases, bug fixes or security updates ³. The announcement in November 2025 gave the reasons: the project had one or two maintainers working in their own time, and flexible features such as snippets annotations, which allow arbitrary NGINX configuration, are now seen as serious security flaws ¹.

Existing installs keep running, and the Helm charts and images stay available ¹². That is the trap: nothing fails on the day, but any vulnerability found later stays open. The Ingress API itself is not being removed, and the separate NGINX controller from F5 is not affected ².

The maintainers point to Gateway API, which Google's open-source team calls "Ingress 2.0", and suggest checking your clusters first, then trying a conversion tool and a conformant implementation such as Envoy Gateway or Cilium ². The check is one command ¹:

kubectl get pods --all-namespaces --selector app.kubernetes.io/name=ingress-nginx

What it means for you: if a tutorial routes traffic with an Ingress NGINX install, it is out of date. Build the lab with Gateway API.

Kubernetes 1.36: quieter, but worth reading

Kubernetes v1.36 shipped in April 2026 with 70 enhancements, according to the release post (its own stage counts add up to 68, and it does not explain the gap) ³. Three items affect everyday work:

• User namespaces are stable. A container's root user maps to an unprivileged user on the host, so an escaped process has no admin power on the node ³.
• The gitRepo volume driver is removed, not just deprecated. Move to init containers or a tool like git-sync ³.
• Service externalIPs is deprecated and warns on use, with removal planned for v1.43. Alternatives are LoadBalancer, NodePort or Gateway API ³.

None of these needs a new course. Each needs a habit: read the release notes for removals before an upgrade, and run the upgrade on a spare cluster first.

What platform engineering 2026 means in practice

The word covers a plain idea: an internal platform team builds paved roads (templates, pipelines, a deployment path) so product teams do not each rebuild them. The 2025 DORA report, based on nearly 5,000 technology professionals, found that 90% of organisations have adopted internal platforms and 76% keep dedicated platform teams, as reported by InfoQ ⁴⁵. Treat these as survey answers about self-reported practice, not an audit.

DORA's more useful finding is about AI. It describes AI as a "mirror and a multiplier": it helps cohesive organisations and exposes the weak spots of fragmented ones ⁴. Good platforms matter because they supply guardrails and shared capabilities ⁵. One oddity: high-quality platforms correlated with a slight rise in delivery instability, which the researchers read as teams with strong recovery experimenting more ⁵.

CNCF's January 2026 survey agrees on the human side. 47% of respondents named cultural change with the development team as the top challenge, and 58% of "cloud native innovators" use GitOps extensively against 23% of "adopters" ⁶. The survey also reports that 82% of container users run Kubernetes in production, up from 66% in 2023 ⁶. CNCF is a foundation that promotes these projects, so read its figures with that in mind. The press release does not state the method; the full report does.

What it means for you: the modern pipeline is not a longer Jenkinsfile. It is a Git repository, a CI run that builds and scans an image, a GitOps controller such as Argo CD that applies the change, and a way to see whether it worked.

AI in the pipeline: two things to separate

AI for writing and reviewing. DORA found 90% of respondents use AI at work, up 14% on the year before, and over 80% say it improved their productivity. Only 24% trust it a great deal or a lot, and 30% trust it little or not at all ⁴. That gap is the skill: you read generated Terraform, YAML and scripts the way you read a stranger's pull request.

AI that works incidents. AWS DevOps Agent became generally available in March 2026. AWS describes it as an always-available operations teammate that triages incidents and correlates telemetry, code and deployment data, now also across Azure and on-premises systems ⁸. AWS says it cuts time to resolution from hours to minutes; that is the vendor's claim, and I found no independent measurement. It is a paid service, with monthly credits for customers on certain AWS Support plans ⁸.

A billing reversal: GitHub Actions

GitHub announced in December 2025 that hosted-runner rates would fall by up to 39% from January 2026, and that self-hosted runners would carry a small per-minute platform charge from March 2026. The rate cut went ahead. GitHub postponed the self-hosted charge to "re-evaluate our approach" and gave no new date ⁷. The page does not say when the postponement was posted.

What it means for you: do not move CI because of a charge that is not in force. Check GitHub's billing page before you plan around either change.

What to skip

• Anything built on Ingress NGINX, however good the video.
• CI migrations driven by the postponed charge, until GitHub says what it will do.
• Tools whose main claim is "AI does DevOps". Judge them by whether you can verify what they did.

What to do with this

Build the lab in this order: Git and CI, a container image, Kubernetes with Gateway API for routing, Argo CD for delivery, then metrics and logs. Then break it, write the runbook, and add one AI step you check by hand. Our earlier note, "Learning DevOps by breaking things: labs, runbooks and a home cluster that teaches", covers the practice loop. TTFA's DevOps Engineer course covers Gateway API, GitOps with Argo CD, platform engineering and AI in incident work, which are the concepts above.

The next note in this series, "CKA or CKAD: what each Kubernetes exam tests and who should sit which", covers what each Kubernetes exam asks you to do and how to choose between them.

Frequently asked questions

What is platform engineering in 2026?

It is building and running an internal platform so developers get a supported path from code to production without rebuilding it per team. DORA's 2025 survey reports 90% of organisations have adopted internal platforms ⁵.

Do I need to learn Gateway API now?

Yes, if you touch Kubernetes. The community Ingress NGINX controller no longer receives security fixes, and the Kubernetes maintainers recommend Gateway API or another supported controller ¹.

Will AI agents replace DevOps engineers?

The sources here do not show that. They show AI used for drafting and incident triage, with low trust in the output ⁴ and vendor claims that need checking ⁸. Verification is the skill that grows.

Is Jenkins still worth learning?

This note's sources do not measure Jenkins use either way. Learn the pipeline ideas first, which carry across Jenkins, GitHub Actions and others.