Skip to content
Technology

KubeCon 2026: agents are now your biggest user

The Deploy Desk 5 min read
KubeCon 2026: agents are now your biggest user

KubeCon + CloudNativeCon North America wrapped up in Salt Lake City last week (November 9–12), and one message came through in almost every keynote: AI agents are no longer a demo. They are running in production, they generate a growing share of traffic, and they are straining the platforms we built for humans.

CNCF’s Jonathan Bryce summed it up as “human speed vs. agent speed.” Below are the numbers behind that phrase, five practical things platform teams should do about it, and the project news worth knowing.

The numbers from the keynote stage

The headline figure: according to Bryce, agents now generate 70% of tracked inference traffic. Here is the rest of what was said on stage and in briefings.

NumberWhat it meansSource
70%Share of tracked inference traffic generated by agentsJonathan Bryce, CNCF
4x in 8 monthsGrowth of inference traffic across providers; network traffic is growing faster than computeCNCF keynote
35%Share of Wikipedia page views driven by agents, which cause most of its resource-heavy trafficJonathan Bryce, CNCF
80%+Share of code written by agents at some Linux Foundation member organizationsJim Zemlin, Linux Foundation
21 millionEstimated cloud native developers, up 36% in a yearCNCF
7 figures/monthNutanix’s AI token bill before it invested $20M in on-prem hardware, expected to break even within a yearDan Ciruli, Nutanix

The pattern is clear: agents make more requests than humans, in less predictable bursts, and the bill arrives faster than most teams expect.

Five things platform teams should do now

1. Authenticate every bot and script

The platforms you depend on are already reacting. On October 19, GitLab.com started enforcing tier-based rate limits, and unauthenticated requests dropped to 60 per hour per IP address. GitLab said it expects load on GitLab.com to grow several-fold during 2026. Premium and Ultimate plans move to the new limits in January 2027.

If any of your CI jobs, scripts or agents call an API anonymously, give them a token now. Authenticated traffic gets higher limits, and it’s the only way to know which automation is causing load.

2. Rate-limit your own APIs per identity, not per IP

What GitLab is doing, you will likely need to do too. Once agents start calling your services, a single customer’s agent can produce more traffic than hundreds of human users. Limit by API key or account, return 429 with a Retry-After header, and publish the limits so agent builders can respect them.

3. Sandbox agents like untrusted code

Bryce said enterprises are slow to adopt agents mainly because they don’t trust autonomous processes. The answer on stage was isolation, using tools most platform teams already know:

  • Kata Containers or gVisor to run each agent in a stronger sandbox than a plain container
  • SPIFFE/SPIRE for short-lived, per-agent identity instead of shared long-lived secrets
  • Envoy-based policy gateways to control which services an agent may call

Solo.io made the enterprise editions of Agent Substrate and kagent generally available for running agents on Kubernetes, and Cosmonic showed WebAssembly-based sandboxes as a cheaper way to run many of them.

4. Start agents read-only, then add approvals

Vendors that let agents act on infrastructure all described the same rollout. Testkube said most customers keep agents read-only or human-in-the-loop when they trigger tests and propose fixes. Akuity, the company behind Argo CD, gives MCP-connected agents read-only or approval-gated permissions. Komodor shows an execution graph before any remediation runs.

A good default: an agent can read logs and open a pull request, but a human merges it.

5. Put a budget on tokens and GPUs

Cost was the quiet theme of the week. Nutanix said 3,000 engineers pushed its monthly token bill into seven figures. CloudBolt now maps GPU, CPU and token costs to individual pods and teams, and rolls out changes in 10% steps with budget caps.

You don’t need a new product to start. Set per-team spending caps with your model provider, tag workloads by team, and review the top consumers every week.

Quick hits: project milestones and launches

  • Graduated projects: the CNCF highlighted OpenTelemetry, KEDA, Kyverno, Cloud Native Buildpacks, CubeFS and KubeVirt as graduated.
  • Newly incubating: ten projects moved up from sandbox, including CloudNativePG, Microcks, Confidential Containers, KubeRay and Open Cluster Management.
  • Postgres on Kubernetes keeps growing: EDB donated its Klio backup and recovery tool to CloudNativePG, and CoreWeave launched a managed PostgreSQL service that runs inside customers’ clusters.
  • New certifications: Kubernetes AI Associate (entry level) and Certified Kubernetes Inference Engineer (hands-on).
  • Uber is open-sourcing Michelangelo, its Kubernetes-native ML platform that serves about 45 million predictions per second.
  • Apple’s container project now creates upstream Kubernetes clusters on Apple silicon via kubeadm, with each container in its own lightweight VM.
  • Security for maintainers: the Linux Foundation launched Akrites, where 30+ enterprises fund engineers to triage and patch upstream vulnerabilities, as AI-generated bug reports flood maintainers.
  • Consistency across distributions: a new Kubernetes Workload Conformance Working Group aims to make apps behave the same on every distribution.

Bottom line

The interesting part of KubeCon 2026 wasn’t a new tool. It was that the answers to “how do we run agents safely?” are mostly things platform teams already do for humans: authentication, rate limits, sandboxing, least privilege, approvals and cost tracking. The difference is speed and volume. Agents will hit every gap in those controls faster than people ever did, so the time to close them is before the traffic arrives.

Sources