▮ agentic-cloud-ops

CNDN 2026/BERGEN/OCT 26–27/120 MIN

Agentic
Cloud
Ops

Practical AI-Driven Kubernetes Operations

From Infrastructure as Code toInfrastructure as Agents. Two hours turning AI agents loose on a live Kubernetes cluster — and figuring out what it actually takes to dare run them in production.

Prepare your laptop →View workshop content

$ git clone https://github.com/leffen/agentic-cloud-ops.git && make preflight # ~10 GB · 15 min · do this at home

lab-1 · agent session · replay● looping
you
Deploy nginx with 3 replicas.
agent
Creating Deployment nginx (replicas: 3).kubectl apply -f nginx.yamlkubectl rollout status deploy/nginx
3/3 ready in 12s.
event
nginx-7c4d-x2 → CrashLoopBackOff (injected by chaos-lab)
agent
Readiness probe hits :8080, container listens on :80. Patching probe port.kubectl logs nginx-7c4d-x2 --previouskubectl patch deploy/nginx --type=json …
Recovered. 3/3 ready. MTTR 41s · human interventions 0.
you
Would you have done that in prod?
agent
No. prod is read-only for me — I would open a PR against the Helm chart and page a human. That is the guardrail.
cluster/k3d-agentic-ops · ns/agentic-opsagent: idle
pods
27
failing
0
agent actions
4
humans on kubectl
0
12:41:28 chaos: readinessProbe.port mutated12:41:35 patch readinessProbe.port 8080→8012:41:42 recovered · mttr=41s12:41:56 policy: prod=read-only · PR drafted
BRING HARDWARE

This is a hands-on workshop. You need a laptop with Docker, ~10 GB of free disk and 8 GB of free RAM — and you must do the setup at home. Conference Wi-Fi is not a build system.

01

Agentic vs. GitOps

What is genuinely new, and what is just marketing. A pipeline executes a sequence; GitOps converges on a declared state; an agent runs a loop. Knowing which you want is the skill.

02

Hands-on experience

Deploy, debug and harden a live cluster with an agent holding kubectl — including the part where it gets it wrong and has to work out why.

03

Risk management

The guardrails that make this safe: namespace-scoped RBAC, tool-layer flags, and a referee that decides what "healthy" means so the agent cannot mark its own homework.

120 minutes

block width = duration · select a block
From IaC to Infrastructure as Agents12 min · talk
Bootstrap and verification8 min · everyone runs this
Buffer5 min · help stragglers
Lab 1 · Your first agent30 min · hands-on
Lab 2 · Security and hardening35 min · hands-on · the centre of gravity
Lab 3 · Chaos and recovery10 min · facilitator demo
Buffer5 min
Q&A and wrap-up15 min
00:55 → 01:30 · 35 min · hands-on · the centre of gravity

Lab 2 · Security and hardening

Start from a deliberately insecure workload and a namespace-scoped RBAC ladder. Harden to restricted, then prove the guardrail holds by watching the agent get refused.

Read the block →
$ make insecure$ make ladder$ make health

Is this
for me?

YES, IF

  • +you use kubectl regularly
  • +you want to watch an agent make real changes to a real cluster
  • +you are curious where the limits actually are
  • +you are a developer, architect or SRE

PROBABLY NOT, IF

  • −you have never seen a Kubernetes manifest
  • −you want an overview presentation to sit back for
  • −you cannot bring a laptop
  • −you expect finished product recommendations

No machine-learning background is needed. None of this is model training.

Prepare before Bergen

A cold bootstrap takes about two minutes on a good connection. Thirty-five of them on conference Wi-Fi takes the whole first block — which is why the prep happens at home, and why this is the only thing we ask of you in advance.

$ git clone https://github.com/leffen/agentic-cloud-ops.git$ cd agentic-cloud-ops$ make preflight
Full preparation guide →
$ make preflightexit 0
OKdocker 27.1runningOKkubectl 1.31foundOKk3d 5.7foundOKpython3 3.12foundOKdisk14.2 GB freeOKimagescached (3.8 GB)6/6checks passed.

What we actually do

Lab 1

Your first agent

Hand over a task in plain language. Watch it plan, act and verify — then debug a deliberately planted ImagePullBackOff without you touching kubectl.

Lab 2

Security and hardening

Harden a deliberately insecure workload to restricted, on a namespace-scoped RBAC ladder. The interesting part is what the agent gets refused.

Lab 3

Chaos and recovery

Incident response end to end, demonstrated at the front of the room with you as backseat drivers. We measure the recovery rather than describing it.

Instructors

Lars Effenberger

TODO: role and one-line bio

Lars Søraas

TODO: role and one-line bio