Incident sheet · /infrastructure
- System
- Infrastructure & Cloud
- Filed
- Read
- 4 min · 737 words
- Severity
- Routine
Agents Build Software; Kubernetes Runs It, KubeCon Data Shows
Zapier data shows only 30% of firms' agents finish most tasks unaided, while China Merchants Bank cut inference costs 60% via Kubernetes — the primitives still decide AI economics.

Event log
- Only 30% of leaders at agent-using organizations say agents complete more than 60% of tasks without human intervention (Zapier, November 9).
- China Merchants Bank raised accelerator utilization from 35% to over 60% and cut inference cost per million tokens by more than 60% using Kubernetes tooling across nearly 10,000 accelerator cards.
- Google announced Agent Substrate 1.0 on November 5 and donated Agent Substrate, gVisor and OpenRL to CNCF.
- Nvidia contributed the OpenShell agent-isolation runtime to the CNCF Sandbox on November 10.
- CNCF and SlashData report the global cloud-native developer population exceeds 21 million; 75% use AI coding tools.
Only 30% of enterprise leaders say their production agents complete more than 60% of assigned tasks without human intervention, according to a Zapier survey published November 9. The finding cuts against the argument that AI agents have made infrastructure primitives obsolete. KubeCon, the flagship Kubernetes (the container-orchestration system for running software at scale) event, drew more than 10,000 attendees last week.
The claim that "no one cares about the primitives anymore" circulated widely as AI enthusiasm spread. Linux Foundation CEO Jim Zemlin rejected it. "Code is cheap [but] trust is expensive," Zemlin said. He summarized the argument in one line: "AI changes how we create software faster than it eliminates the work of operating it."
How much of the agent hype is real?
The Zapier survey polled 513 AI-experienced US managers at companies with at least 1,000 employees. It found 71% of respondents' organizations had agents in production. That is genuine adoption, but far from autonomous operation of the business.
A second dataset complicates the picture further. A CNCF and SlashData report found the global cloud-native developer population has surpassed 21 million. Of those, 75% use AI coding tools, compared with 55% of non-cloud-native developers. The Kubernetes infrastructure community shows no signs of resisting AI; the two interests overlap.
What did China Merchants Bank achieve?
A September CNCF case study documented the strongest numbers in the story. China Merchants Bank ran AI training and inference across nearly 10,000 accelerator cards. Its problem was not hardware supply. It was predictability: training requires predictable capacity, while customer-facing inference must respond to changing demand.
The bank combined Kubernetes with tooling for scheduling, monitoring, scaling and model-data movement. The results were concrete:
- Average accelerator compute utilization rose from 35% to more than 60%.
- Inference cost per million tokens fell by over 60% for comparable models and services.
The primitives, in other words, cut the cost of AI dramatically.
What did Google and Nvidia contribute?
New AI workloads are generating new infrastructure requirements. On November 5, Google announced Agent Substrate 1.0, generally available on Google Kubernetes Engine, and donated Agent Substrate, gVisor (a container isolation layer) and OpenRL to the CNCF.
Google's design addresses a familiar problem set. Agents execute code, wait for model responses, retain state and need access limits. Holding dedicated compute while an agent waits wastes money; running generated code without adequate isolation creates risk. Google's approach separates persistent agent state from reusable execution environments and enforces access through gateways.
Nvidia acted on November 10, contributing OpenShell to the CNCF Sandbox. The runtime enforces boundaries around agents' access to data, secrets and networks. Zemlin endorsed the move: OpenShell is "critical sandbox software that needs peer review and collective development."
The open-source contribution model matters here. Companies can share the work of testing a security boundary rather than each trusting a fresh implementation an agent produced. Open code does not guarantee safety; it provides a common place to scrutinize the controls.
Who owns agent failures?
Kubernetes creators Craig McLuckie and Joe Beda framed the accountability question in an interview earlier this year. Enterprises own the outcomes of work they delegate to agents, regardless of who or what performed it.
That ownership has a practical shape: permissions enforced, state preserved, resources allocated and failures contained. Even if agents eventually handle most configuration and troubleshooting, the platform determines what they can accomplish. Stable interfaces, enforced permissions and accurate feedback make automation possible; an undocumented pile of scripts does not become a reliable platform because a machine now has access to it.
Isovalent cofounder Thomas Graf made the adjacent point in 2023. Platform teams must meet enterprise requirements "without disrupting and undoing the gains that have been made on agility and developer efficiency."
Does every project need Kubernetes?
No. A managed service that hides the infrastructure may be the better answer in some cases. The operative principle is reuse: deploy capabilities somebody knows how to operate rather than generating another pile of infrastructure because an agent can produce it.
The recommended baseline for agent platforms includes enforced permissions, spending limits and a mechanism to undo a bad deployment. Asking an agent to generate an application and then reassure its operators that everything is fine does not constitute an operating strategy. Zemlin's formulation stands: code is cheap, trust is expensive, and someone still has to care about the primitives.
via infoworld.com/feed/ (Source)
More from Rebecca Stone
Show full bio
Senior reporter covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.
199 articles
The newsroom digest
One email a day with the stories that matter. No noise.
Cross-references
- Accenture and Google Cloud Field 1,000 Engineers for Enterprise AI
- KubeCon 2026: Three Rules for Enterprises Running AI Agents
- Nvidia and Google bring agent sandboxes under CNCF governance
- Two-thirds of enterprises now run AI agents: reading the 68% figure
- Microsoft, Google Move AI Agent Governance Into Enterprise IT Mainstream