Skip to content
Let's talk

Futurum Research

Insight · Events

AI Platforms8 min read

AI Agents Gain an Open Runtime as Agent Governance Lags at KubeCon + CloudNativeCon NA 2026

CNCF Adds Agent Runtimes From Google and NVIDIA as Solo.io and Red Hat Ship Agent Controls, but Identity and Approval Still Lag

Research Director, Software Lifecycle Engineering

The short answer

Vikram Rathnam and Mitch Ashley of The Futurum Group share their insights on KubeCon 2026, where AI agents gained open runtimes on Kubernetes and agent governance became the limiting factor in deploying them.

Futurum's Vikram Rathnam and Mitch Ashley,

Cite this

Vikram Rathnam and Mitch Ashley, The Futurum Group, "AI Agents Gain an Open Runtime as Agent Governance Lags at KubeCon + CloudNativeCon NA 2026," November 17, 2026. https://trial.futurumgroup.com/insights/ai-agents-gain-an-open-runtime-as-agent-governance-lags-at-kubecon-cloudnativecon-na-2026/

AI Agents Gain an Open Runtime as Agent Governance Lags at KubeCon + CloudNativeCon NA 2026

What is Covered in this Article

  • KubeCon 2026 in Salt Lake City and its new AI Inference + Agentic track
  • Agent runtimes joining CNCF: Agent Substrate, gVisor, and OpenShell
  • Agent controls from Solo.io, Red Hat, SUSE, Cloudsmith, and the Linux Foundation
  • Why agent governance now limits AI agent deployment

The Event—Major Themes & Vendor Moves: KubeCon 2026, the annual North American edition of KubeCon + CloudNativeCon, ran November 9–12 in Salt Lake City. The event site lists more than 9,000 attendees and over 400 sessions. The Cloud Native Computing Foundation (CNCF), which hosts Kubernetes, added an AI Inference + Agentic track this year, covering agent orchestration and model serving.

On November 10, CNCF announced new projects, including Agent Substrate, a zero-trust runtime for artificial intelligence (AI) agents on Kubernetes; gVisor, a user-space kernel that isolates workloads; and NVIDIA's OpenShell, a runtime for autonomous agents. Google is donating Agent Substrate and gVisor. CNCF also said its Kubernetes AI Conformance Program reached 40 certified platforms. The llm-d project, for distributed large language model (LLM) inference, shipped 1.0 on October 2. Broadcom contributed a VM API to bring virtual machine provisioning and management under Kubernetes APIs as AI and agentic workloads pull VMs into cloud-native environments, and it co-chairs a newly formed Workload Conformance Working Group.0, ahead of the event.

Vendors added controls on top. Solo.io made its enterprise Agent Substrate and kagent generally available, with agentgateway enforcing identity and policy on agent traffic. Red Hat OpenShift 5 ships a Model Context Protocol (MCP) server that defaults to read-only access and tags actions for audit logging, plus Agent Sandbox in Technology Preview. SUSE made its Virtualization MCP server generally available, with a person approving every action. Cloudsmith launched an AI Skills registry in Early Access that allows or blocks agent skills by version. The Linux Foundation launched Agent Name Service, which uses the Domain Name System (DNS) to give agents. Observability vendors worked on the same gap through OpenTelemetry. Dynatrace contributed to OpenTelemetry at the event and presented on benchmarking open-source agent memory on Kubernetes with it, the kind of open, portable record an audit can read for verifiable identities.

Analyst Take: KubeCon landed differently depending on attendees’ roles. For practitioners, agents stopped being a side experiment and gained real runtimes and controls to build on, under CNCF's vendor-neutral stewardship. For IT leaders, the same week moved agent governance and accountability from a research topic to a buying and risk decision.

KubeCon also opened the door for vendors to extend their product strategies into the new agent layer, where runtimes, controls, identity, and the connector surface remain unclaimed. NVIDIA's OpenShell stood out as the lead open-source contribution, a runtime for securing and governing autonomous agents, with hundreds of organizations already using it, tens of thousands of contributions in under a year, and NVIDIA and Google aligning it with Google's Agent Substrate.

KubeCon 2026 Made Agent Governance the Deployment Bottleneck

KubeCon 2026 showed that AI agents will run on Kubernetes, in open runtimes under CNCF's vendor-neutral stewardship. Agent Substrate, gVisor, and OpenShell enter at Sandbox, CNCF's earliest stage, but the direction is set. What limits agent deployment now is agent governance, meaning what operators can control and prove, not what Kubernetes can run. AI conformance certifies the platform and says nothing about what an agent does on it.

Deployment is already ahead of control. CNCF and SlashData found that 49% of developers who build AI into their services report having agentic systems in production. In Futurum's 2H 2026 Software Lifecycle Engineering Decision Maker Survey of 839 IT decision makers (July 2026), 75% of organizations reported a production incident in the past 12 months in which AI was a contributing factor. Only 18% have standardized or mastered AI agent governance, the least mature of the eight practices we measured.

Today's Agent Controls Limit What an Agent Can Reach

The KubeCon 2026 announcements from Solo.io, Red Hat, Cloudsmith, and the Linux Foundation put controls at the edge of the agent: sandboxes, gateways, read-only defaults, and allow lists that limit what an agent can access. A site reliability engineering (SRE) or security team also has to know which agent acted and who authorized it. SUSE's approve-every-action design answers who authorized a change; teams also need that approval kept with the change afterward.

Agent platform vendors should compete on what their runtime records, since the CNCF is making isolation a shared default. Before production, platform and SRE teams should test agent governance by asking three things of any agent runtime: does each action carry the agent's own identity, does the approval travel with the change, and can the record leave the vendor's tool in an open format such as OpenTelemetry?

Buyers Leave KubeCon With More Control Planes to Govern, Not Fewer

KubeCon 2026 answered the platform team's question and broadened the buyers'. Whether agents run on Kubernetes is settled: they will, in open runtimes under CNCF stewardship. What the enterprise now owns is everything stacked on top. In one week, Solo.io, Red Hat, SUSE, Cloudsmith, and the Linux Foundation each shipped a piece of agent control, and few of those pieces were built to work together. An enterprise that adopts several inherits the integration, governance, and cost debt between them. That is control plane debt, and it accrues fastest in exactly the moment the market is in now, before the layer roles settle and a dominant few emerge.

The decision no longer sits with the platform team alone. Agent governance spans the CISO, who answers for what an agent did and who authorized it; the CIO, who governs the connector layer agents reach through; and the CFO, who owns a runtime cost metered by consumption rather than by seat. The durable leverage in that group is authorizing what an agent can reach, not owning the runtime it runs on. Most of the week's controls sit at that layer: Red Hat's MCP server defaulting to read-only, Solo.io's agentgateway enforcing identity for agent traffic, and the Linux Foundation's Agent Name Service providing agents with verifiable identities via DNS. The buyer who governs their scales agentic deployment. The buyer who governs application by application stalls and pays for the gap in audit findings later. MCP is becoming the connector every vendor builds to, keeping the connector layer the place where the CIO still sets the terms.

Governance maturity, not runtime availability, is the constraint that limits deployment, and the buyer-side evidence is blunt. Futurum's 2H 2026 Software Lifecycle Engineering Decision Maker Survey found that 75% of organizations reported a production incident in the past year in which AI was a contributing factor, and only 18% reported standardized agent governance. Accountability for that gap surfaces in incident response and audit before it reaches procurement, which is the wrong order. The three tests this analysis applies to any agent runtime, identity on every action, approval that travels with the change, and a record that exports in an open format, belong in the contract and the RFP, not only the architecture review. The economics belong there too. A runtime metered by consumption makes agent governance as much a cost-control question as a risk one, and a buyer who cannot see the cost per agent action cannot budget for the deployment just approved.

You can read the full CNCF announcement on the CNCF website.

What to Watch

  • Agent Substrate Adoption: If distributions beyond Google Kubernetes Engine and Solo.io's enterprise edition ship it, Agent Substrate could become the common agent runtime.
  • Red Hat Agent Sandbox: Watch for general availability and whether audit tagging extends from the MCP server to actions inside the sandbox.
  • Agent Identity Standards: Agent Name Service needs Kubernetes platform vendors to adopt it, or each runtime will keep its own identity model.
  • Control Plane Consolidation: Watch whether buyers consolidate onto a few agent control planes or accumulate one per vendor. The enterprises that consolidate pay down control plane debt. The rest carry it into the next budget cycle.

Sources

1. Observability: AI Is Eating the Budget That Keeps AI Safe, Futurumgroup

2. Claude Agent Observability Can't Be the Audit Record on Its Own, Futurumgroup

3. Docker Reruns the Container Playbook, This Time for Cloud Sandbox Kit, Futurumgroup



Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.

Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.

Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.

Read the full Futurum Group Disclosure.

Published by Futurum.

About the author

Vikram Rathnam, Research Director, Software Lifecycle Engineering

Vikram Rathnam

Research Director, Software Lifecycle Engineering

Vikram Rathnam is Research Director for Software Lifecycle Engineering at The Futurum Group. His coverage follows the software lifecycle as AI agents take on more of the work: development, testing, security, operations, platform engineering, observability, the control planes that govern agents, and the open standards that let them interoperate. He studies who owns the outcome when an agent acts and what evidence shows whether the agent did what it was supposed to do. He came to analyst work from the vendor side. Over 24 years, he worked first as an engineer, including at Silicon Labs, and later in product and partner strategy at Cisco, Secureworks, Dell, and CMIT Solutions. Across those roles, he managed B2B cybersecurity, IoT, and AI products and worked extensively with partner and channel ecosystems. At CMIT Solutions, he led product strategy across a franchise network of 180 independently owned locations. Franchisees ranged from shops with dedicated technical teams to two-person offices. He tests vendor positioning against the technology, the practitioner who has to adopt it, and the partner who has to sell it. Vendors briefing him get an analyst who has participated in roadmap reviews, pricing debates, and channel enablement meetings and can identify where a strategy will work and where it will fail. Expertise: Product strategy, partner and channel go-to-market, cybersecurity (XDR and managed security), AI-native product development, software lifecycle engineering, and observability Education: MS, Michigan Technological University; MBA, The University of Texas at Austin Location: Austin, Texas

Coverage on Alphabet (Google/Google Cloud)
  • Analyst Insights3
All Alphabet (Google/Google Cloud) coverage
Coverage on Cloud Native Computing Foundation (CNCF)
  • Analyst Insights2
All Cloud Native Computing Foundation (CNCF) coverage
See the full picture in the platform

Vendor scoring, the benchmark data behind these findings, and the analyst team that produced them.

Open the platform (opens in a new tab)Become a client
The Weekly

Subscribe to The Weekly

The week's market signals, one email, no filler. Unsubscribe any time.