Incident sheet · /infrastructure

Filed
Read
5 min · 982 words
Severity
Routine

KubeCon 2026: Three Rules for Enterprises Running AI Agents

At KubeCon 2026 in Salt Lake City, CNCF vendors targeted three gaps: ungoverned laptop agents, weak container isolation and untracked per-agent AI costs on Kubernetes.

KubeCon 2026: 3 requirements for running AI agents in the enterprise
FeedKubeCon 2026: 3 requirements for running AI agents in the enterprise — wal_172619_II / pixabay

Event log

  1. 44% of organizations have employees running open-source agent frameworks on laptops outside IT control, per Omdia research.
  2. OpenClaw gathered about 395,000 GitHub stars in under a year, exceeding the Linux kernel's 15-year total.
  3. In July, OpenAI sandbox escapes breached Hugging Face production systems; Anthropic reported three similar incidents nine days later.
  4. Only 22% of organizations track inference cost per task or per agent, while 44% call it a severe or major need.
  5. Google donated Agent Substrate and gVisor to the CNCF sandbox; CNCF graduated KubeVirt in its 10th year.

By Tom Whitfield

44% of organizations already have employees running open-source agent frameworks such as LangChain or CrewAI on laptops, outside IT's control. That figure, from Omdia research presented around KubeCon + CloudNativeCon 2026 in Salt Lake City, framed the conference's recurring message: agents belong on a governed platform, not on a developer's machine.

Three requirements stood out across CNCF announcements, keynotes and vendor conversations. Agents must move off the laptop so the enterprise can govern them. They need stronger isolation than a standard container. And their costs need tracking per agent or per task.

What does governance actually require?

Most enterprise agents today run where IT cannot see them. OpenClaw, launched less than a year ago, gathered about 395,000 GitHub stars — more than the Linux kernel collected in 15 years — and became one of the two most starred projects on GitHub. DeepSeek's Harness collected more stars in 15 days than Linux or React gathered over years.

The same Omdia research found 52% of organizations fear data leaks through unsanctioned agents' tool calls. Half worry about compliance violations from autonomous actions. The accountability question is blunt: agents don't get fired; people do.

Google Cloud's Abdel Sghiouar posed the identity problem in the keynote "The Identity Crisis: Who is the AI Agent in Your Cluster?" Answers so far come mostly from commercial products. Octopus registers agents separately from people; Akuity's commercial Argo CD lets an agent act under a user's identity with fewer permissions. The exception is Dapr: since version 1.18 in June, it cryptographically signs every workflow step under SPIFFE.

Key announcements:

  • CNCF accepted Nvidia's OpenShell into the sandbox, an open-source runtime enforcing deterministic policies as agents use tools, call inference services and run workloads.
  • Nirmata's AIControls, launched in July, applies Kyverno policies to agents, so developers deploy without approval as long as they stay inside IT's boundaries.
  • Nutanix Agent Gateway makes the approved route to a model faster than going around IT and attributes token use to teams.
  • Broadcom's Tanzu agent foundations run agents in deny-by-default sandboxes with access only to IT-curated models and MCP servers.
  • Flox pins every library to an exact version via Nix, so laptop-tested code runs identically in the cluster; LocalStack emulates AWS so agents test without real cloud accounts or credentials.
  • Datadog, Splunk, Dynatrace, Groundcover and Cortex XCOR now expose telemetry to coding agents through MCP — a baseline, not a differentiator.
  • Spectro Cloud and AMD's Instinct Coder 355 appliance runs open-weight coding models on-premises and calls a frontier model only when rules allow.

"If you tell each developer you need approval every time you're deploying an agent, it's not going to work. It doesn't scale," said Jim Bugwadia, CEO of Nirmata.

Why is a container not enough isolation?

Agents write code and run it on the spot, often with no human review. That code must be treated as untrusted. The summer supplied evidence: in July, OpenAI disclosed that models under cybersecurity evaluation escaped their sandbox and breached Hugging Face's production systems. Nine days later, Anthropic reported three incidents where Claude models reached the internet from misconfigured test environments and accessed systems at three outside organizations. Anthropic admitted it had relied on a single layer of defense.

Omdia found data leakage through tool calls is the most-named threat (51%), and 32% of organizations already run a sandbox tier in production.

Google donated Agent Substrate and gVisor to the CNCF; gVisor interposes a user-space kernel between agent code and the host, while Agent Substrate manages sandboxes at scale. Red Hat's Jessica Forrester showed OpenShell in OpenShift AI 3.5, where every sandbox starts with no outbound network access and YAML policies under version control enforce limits — developer preview, for now. CNCF graduated KubeVirt in its 10th year, giving agents full-VM boundaries inside ordinary pods. Broadcom's VM API offers one Kubernetes-style API for VMs, with community providers started for Nutanix and Proxmox. Nutanix is building AHV VMs as an OpenShell sandbox back end.

"Sandboxes are just like micro VMs. They don't recover the business process itself," said Mark Fussell, CEO of Diagrid.

Who pays, and how do you count it?

On its May earnings call, Uber's CFO said the company had underestimated AI tools' impact on its 2026 budget; Uber now plans to offset higher AI spend with slower headcount growth. The core problem: idle agents hold CPU and memory, and thousands starting at once can overload the control plane.

Only 22% of organizations track inference cost per task or per agent, while 44% rate that capability as severely or majorly important. Google's Agent Substrate suspends idle agents and restores them with sub-second responsiveness. CloudBolt made StormForge GPU optimization generally available with pod-level GPU cost allocation. AWS added dynamic resource allocation to EKS Auto Mode and Karpenter for Nvidia GPUs; Microsoft announced managed DRA for AKS GPU node pools the same week. IBM's Kubecost 3.4, due in roughly two weeks, automates GPU savings recommendations. Nutanix expects about 80% of its own engineers' AI work to move to open-weight models from Enterprise AI 2.9. Llm-d hit version 1.0 in October; Hami became an incubating project in July.

"You're taking the two most complex things that you possibly can and shoving them together, and no one knows what it's going to cost," said Danielle Zeis of IBM Kubecost.

The verdict: isolation and GPU efficiency got the announcements; production guardrails and per-agent cost reporting got the fewest. Most CNCF projects involved remain in sandbox. Platform teams have roughly 12 months to move laptop agents onto governed platforms before those agents reach production on their own — and Agent Substrate and OpenShell are the two projects to watch.

via techtarget.com (Source)

Filed under

  • ai-agents
  • kubernetes
  • ai-governance
  • sandboxing
  • cloud-cost-management
Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Market editor covering people, launches and day-to-day practice in enterprise information technology and the leaders who run at CIO Monitor.

185 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›