DevOps Watch: Kubernetes, AI Agents and the New DevOps Control Plane
Kubernetes is hardening its role as the infrastructure foundation for AI and cloud-native workloads, while AI agents are moving deeper into software delivery. Across platform engineering, CI/CD, and security, the DevOps control plane is beginning to manage not only human developers but also autonomous software workers.
The traditional CI/CD pipeline is becoming an increasingly automated software-delivery system. Kubernetes 1.37 strengthens the platform underneath modern cloud-native workloads, while Atlassian, Harness and Anthropic are developing systems in which AI agents can plan, write, test and review software. The operational challenge is to increase automation without losing human approval, auditability, security, and control.
Kubernetes 1.37 Strengthens the Cloud-Native Foundation
KUBERNETESKubernetes 1.37, known as “Garhwal,” continues the project’s steady movement toward a more mature production platform. The release contains 67 enhancements, including 16 general-availability milestones, 23 beta features, 27 alpha features,s and one deprecation.
One of the most important changes is the graduation of the Kubernetes Metrics API to stable status. The API provides a standard way to retrieve CPU and memory usage for pods and nodes, supporting tools such as kubectl top and the Horizontal Pod Autoscaler.
Metrics are more than an observability feature. They influence scaling decisions, capacity planning, and incident diagnosis. A stable API gives platform teams greater confidence when building autoscaling systems and internal developer platforms around Kubernetes resource data.
Kubernetes 1.37 also promotes KubeletInUserNamespace, commonly described as Rootless Kubelet, to beta. The feature allows core node components to run as non-root users through Linux user namespaces, reducing the potential impact of some node-level security failures.
Platform-team implication: Rootless Kubelet should be tested in controlled node pools before broad adoption. Teams need to verify compatibility with container runtimes, networking, storage plugins, monitoring agents, and host-level automation.
AI Agents Enter the Software-Delivery Loop
AI-NATIVE DEVOPSAtlassian’s current platform direction illustrates the transition from AI-assisted coding to AI-assisted software delivery. Its “governed agent loops” connect work planning, coding, testing,g and review while keeping policy and approval controls inside the workflow.
A governed delivery loop can include:
- Planning and issue decomposition.
- Code generation and repository updates.
- Automated testing and validation.
- Code review against project standards.
- Deployment preparation and approval.
- Incident investigation and remediation proposals.
The important control question is not whether an agent can complete a task. It is whether the organization can define which actions require human approval, which environments the agent can access, ss and how every decision is recorded.
The emerging pattern is therefore not unrestricted autonomy. It is automation surrounded by identity, permissions, testing, auditability and operational boundaries.
Harness Designs a Control Plane for Agentic Engineering
DEVOPS PLATFORMSHarness has previewed a redesigned platform and its forthcoming Software Factory for engineering teams managing AI-assisted and agentic workflows. The platform is designed around the idea that organizations may soon need to coordinate teams of software-engineering agents rather than a single coding assistant.
That creates a new platform-engineering requirement: the DevOps control plane must represent non-human workers as operational identities.
| Control area | Operational question |
| Identity | Which agent or workflow performed the action? |
| Permissions | Which repositories, environments,nts and tools can it access? |
| Testing | Which checks must pass before the agent’s work advances? |
| Approval | Which changes require a human decision? |
| Auditability | Can the organization reconstruct what happened? |
This is why agentic DevOps is not simply a developer-productivity feature. It is a new control-plane problem involving access management, policy enforcement, observability,lity and software-delivery governance.
Claude Code Preserves Branches and Pull Requests
AI CODING INFRASTRUCTUREAnthropic’s redesigned Claude Code Projects experience provides another example of AI entering the delivery workflow. A coordinator can divide a larger development objective among parallel Claude Code sessions, with each worker operating on an individual branch and repository copy.
The significance for DevOps is the preservation of familiar software-delivery primitives. Branches, pull requests,uests and merge conflicts remain part of the workflow even when agents perform the underlying implementation work.
That approach creates a bridge between experimental agentic development and established engineering controls. Organizations do not need to replace source control or review systems immediately. Instead, they can use those systems to constrain and inspect increasingly autonomous work.
The new bottleneck may be review capacity. If several agents can work in parallel, teams must ensure that automated tests, security checks, and human reviewers can evaluate the resulting increase in change volume.
GitHub Actions Runner Enforcement Nears
CI/CD OPERATIONSGitHub Enterprise Cloud will begin full enforcement of its minimum version requirements for self-hosted GitHub Actions runners on September 25, 2026. Runners must be version 2.329.0 or later to register, while continued job execution requires runners to remain within the applicable rolling update window.
What teams should check now:
- Inventory every self-hosted runner, including ephemeral and autoscaled fleets.
- Confirm that runners meet version 2.329.0 or later.
- Verify that update automation installs new runner versions within 30 days.
- Test registration and job execution during brownout periods.
- Check custom labels, local tools, and persistent workspace dependencies.
- Confirm that replacement runners can be provisioned quickly.
- Review runner tokens, cloud credentials and outbound network permissions.
This deadline is more than a version-management task. Self-hosted runners are privileged execution environments connected to source repositories, package registries, cloud accounts and deployment systems. Runner lifecycle management is therefore part of CI/CD security.
Software Supply-Chain Security
DEVSECOPSThe 2026 Trivy-related supply-chain campaign remains a significant DevSecOps lesson. Reporting described how an attack against Aqua Security’s Trivy security scanner and associated GitHub Actions expanded to other development tools, including Checkmarx’s KICS and LiteLLM.
Later reporting cited CloudSEK estimates that more than 2,500 companies and approximately 434,000 CI/CD pipelines were affected by the broader campaign.
The relevant security chain is:
Source code → dependencies → CI/CD actions → containers → registries → deployment infrastructure
Every link should be treated as a security boundary. Practical controls include pinning third-party actions to immutable commit SHAs, limiting token permissions, isolating runners, separating build and deployment credentials, signing artifacts, and monitoring unusual outbound traffic from CI jobs.
Platform Engineering Becomes the AI Control Plane
PLATFORM ENGINEERINGInternal developer platforms traditionally provide templates, deployment workflows, environment provisioning and operational visibility. The AI-native platform must add another layer: it must determine what agents know, what they can do, and how their work is measured.
A platform may need to provide:
- Context controls for repositories, tickets, documentation,n and incidents.
- Identity and authorization for each agent and workflow.
- Sandboxed environments for generated code.
- Standardized development and deployment policies.
- Approval gates for production and infrastructure changes.
- Observability for agent activity, cost, latency,cy and failures.
- Reproducible records of prompts, tools, artifacts and decisions.
The platform team’s role is therefore expanding from creating paved roads for developers to creating governed operating environments for both developers and autonomous software agents.
What DevOps Teams Should Watch
STRATEGIC PRIORITIES- Kubernetes 1.37 adoption — Metrics API compatibility, Rootless Kubelet pilots and production upgrade planning.
- GitHub Actions runners — Version compliance, update automation and fleet recovery before September 25.
- Agent permissions — Repository, cloud, deployment and production-access boundaries.
- AI-generated changes — Pull-request review capacity, testing depth and rollback procedures.
- Supply-chain controls — Action pinning, artifact signing, runner isolation and token minimization.
- AI-native platforms — Identity, observability, cost control and governance for agent fleets.
- Incident automation — AI-assisted investigation with human approval for remediation.
DevOps Watch at a Glance
| Area | Development | Operational significance |
| Kubernetes | Version 1.37 | Stronger platform foundation for cloud-native and AI workloads. |
| AI delivery | Governed agent loops | Agents participate across planning, coding, testing,g and review. |
| CI/CD | Runner enforcement | Self-hosted fleets require immediate version and lifecycle checks. |
| DevSecOps | Supply-chain campaigns | CI/CD tools and actions are critical security boundaries. |
DevOps is becoming a system for coordinating humans, agents, entities, and infrastructure. Kubernetes remains the foundation, while AI agents increasingly participate in planning, implementation, testing, and review.
Automation does not remove the need for control. Identity, permissions, approval gates, auditability and supply-chain security become more important as agents receive broader access.
The next internal developer platform will manage agents as well as developers. The strongest teams will be those that increase delivery speed without surrendering operational visibility or security boundaries.
Source Attribution
- Kubernetes — “Kubernetes v1.37: Garhwal,” August 26, 2026.
- InfoQ — “Kubernetes 1.37 Released: Stable Metrics API and Rootless Kubelet,” September 2026.
- Atlassian — “We’re Bringing Governed Agent Loops to the AI-Native SDLC,” September 2026.
- DevOps.com — “Harness Previews Revamped Platform for the Agentic Engineering Era,” September 2026.
- DevOps.com — “Anthropic Adds a Coordinator to Claude Projects for Running AI Work in Parallel,” September 2026.
- GitHub Changelog — “GitHub Actions: Minimum Version Enforcement Timeline for Self-Hosted Runners,” June 2026.
- DevOps.com — “Sophisticated Supply Chain Attack Targeting Trivy Expands to Checkmarx, LiteLLM,” March 2026.
- DevOps.com — “LiteLLM Attack Affected 2,500 Companies, 434,000 CI/CD Pipelines,” August 2026.
- InfoQ — Platform-engineering coverage on AI agents and internal developer platforms, 2026.
THE CODEW · DEVOPS WATCH
Editorial Note
DevOps Watch tracks how engineering teams build, test, deploy, secure, and operate software. Coverage focuses on delivery systems, cloud-native infrastructure, platform engineering, CI/CD and the operational control layers surrounding modern development.
This coverage is based on public disclosures and reporting cited above and should be read in the context of those sources and their reporting period.
Reviewed by Erwin Castro
on
Tuesday, September 22, 2026
Rating:
