Welcome to the NZRT Wiki Podcast. Today we’re looking at Kagent Setup — NZRT.

This episode walks you through installing kagent in the NZRT Kubernetes cluster and wiring it up to use Claude Sonnet as the primary language model. By the end you’ll have a working SRE agent running inside the cluster and ready to diagnose problems.

Let’s start with what you need before you begin. There are five prerequisites. First, a Kubernetes cluster — NZRT already has prod, staging, and dev namespaces set up. Second, kubectl configured with cluster access. Third, Helm version 3 for the chart installation. Fourth, an Anthropic API key from the NZRT Gmail Anthropic account. And fifth, the kagent CLI itself, which we install in the very first step.

Step one is getting the kagent CLI onto your machine. You run a one-line install script that downloads and runs the installer automatically from the kagent website. Once that finishes, you check that it worked by asking for the version number. If a version comes back, you’re good to move on.

Step two is creating the kagent namespace in your cluster. A single kubectl create namespace command handles that. You then label the namespace with three NZRT standard values — owner set to nzrt, service-code set to 000K8S, and environment set to prod. Those labels keep things consistent with how NZRT organises all its cluster resources.

Step three is storing your Anthropic API key as a Kubernetes secret. You create a secret called anthropic-key inside the kagent namespace, passing your API key in as a literal value. You then run a get secret command to verify it’s there. Once you see it listed, the key is safely stored and ready to use.

Step four is the main Helm installation. You add the kagent chart repository, update your local cache, then run the install command. That command puts kagent into the kagent namespace and you pass in three settings: the default provider is Anthropic, the secret name is the one you just created, and you specify which field inside that secret holds the actual key value. After the install, you check the rollout status of the kagent controller deployment and list all pods in the namespace to confirm everything is running.

Step five is creating a ModelConfig resource. This is how you tell kagent exactly which Claude model to use. You write a small YAML file defining a resource of kind ModelConfig, name it claude-sonnet, place it in the kagent namespace, and add the NZRT owner and service-code labels. In the configuration spec you reference the anthropic-key secret, set the model to claude-sonnet-4-6, and set the provider to anthropic. You apply the file with kubectl and the model config becomes live. The wiki also notes that for heavier reasoning tasks you can create a second ModelConfig pointing at claude-opus-4-7.

Step six is deploying the NZRT SRE agent, and this is where everything comes together. The agent is defined in another YAML file, this time as a resource of kind Agent. It’s named nzrt-sre-agent, lives in the kagent namespace, and carries the standard NZRT labels. The most important part is its system message — the instruction set that defines what the agent does and how it behaves. It tells the agent it is an SRE agent for the NZRT cluster, lists the namespaces it works across — prod, staging, dev, monitoring, and kagent — and sets firm boundaries. It must use read-only tools first, it must never delete or modify resources, and when it has a fix recommendation it should output the exact command or YAML but wait for human confirmation before anything gets applied. The agent is linked to the claude-sonnet ModelConfig from the previous step. It also gets five tools from the kagent tool server: get resources, get pod logs, describe a resource, get cluster events, and check service connectivity. You apply the file with kubectl just as before.

Step seven is opening the dashboard. A single kagent dashboard command launches a local web interface at port 8082. If you prefer to do it manually, you can use kubectl to port-forward the kagent dashboard service to that same port on your machine.

To know that everything is working correctly, run through this mental checklist. All pods in the kagent namespace should show as Running. Your claude-sonnet ModelConfig and your nzrt-sre-agent should both show a Ready status. The dashboard should be reachable at localhost on port 8082. And when you send the agent a test message asking it to list all pods in nzrt-prod, it should respond with that pod list — confirming the model connection and tool access are both functioning end to end.

That’s it for this episode of the NZRT Wiki Podcast. Thanks for listening.