Install kagent 1.0 on a local Kubernetes cluster
Code Snippets
install:
kagent:
valuesFiles: [kagent-values.yaml]config:
server:
grpcReflection: trueInstall kagent 1.0 on a local Kubernetes cluster. In this video you'll install kagent 1.0, with Agent Substrate, on a Kubernetes cluster that runs on your laptop. kind runs a whole Kubernetes cluster inside one Docker container on your laptop. On that cluster runs Agent Substrate, the layer that runs every agent in its own sandbox, and on top of it kagent: the controller, the web interface, the tool server and a database. In this lesson you will create a Kubernetes cluster with kind, install kagent with its command line tool, and look at what it runs. Before you start, you need Docker, kubectl, Helm version 3 and kind. The first commands check all four. Docker, kubectl, Helm version 3, and kind, which runs Kubernetes inside Docker containers. kagent 1.0 is still a release candidate, so I pin one exact version, 1.0.0 release candidate 3, in every command. kagent 1.0 needs Kubernetes 1.37 or later. Agent Substrate depends on two certificate APIs that became generally available in that release. kind can create a cluster at that version if I give it the matching node image. kind starts one container that acts as the whole cluster. It takes about a minute, so I speed up the video. Inside that container run the Kubernetes control plane, which decides what runs where, and a node that runs the workloads. The checklist shows kind installing a network plugin and a storage class. Let me check that this cluster serves the two certificate APIs that Agent Substrate needs: PodCertificateRequest, and ClusterTrustBundle. Now the kagent command line tool. I download the installer script and ask for exactly version 1.0.0 release candidate 3. I add the no-sudo flag, so it installs into my home directory instead of a system folder. The installer takes its settings from two small files. The first is a Helm values file. It turns on gRPC server reflection in the controller, which makes the controller's API discoverable by tools like grpcurl. The second file is the installer's own configuration. Its only job is to point at the values file. Now I run the installer. Its default provider wants an API key in the environment, so I exported a placeholder before recording. My agents do not use that provider: each agent names its own model endpoint. This takes about three minutes, so again I speed up the video, this time seven times. The installer works in order. First the custom resource definitions, which teach Kubernetes new kinds of objects, such as an agent and a session. Then Agent Substrate, which runs each agent as a sandboxed actor. Then a database for sessions and conversation history. Last, kagent itself: the controller, the web interface, and the tool server. Installed. Let me look at what is running. The kagent namespace holds the controller, the user interface, the tool server, the database, and a pod called kagent-default. That one is a worker, a pod that agents run on. In the ate-system namespace live the Agent Substrate components: the API server, the identity server, the egress gateway and the router, the node agent called atelet, and an object store for snapshots. A worker pool called kagent-default is ready, with one worker. That is the capacity my agents will run on, and each agent runs inside a gVisor sandbox on it. kagent also has a web interface. I forwarded its port before recording, so I can open it on localhost, port 18080. The dashboard counts what is installed: agents, model configurations, MCP servers, and the tools those servers offer. There is no agent yet. The tools come from the two MCP servers that kagent installed itself. The Substrate page shows Agent Substrate. Here is the worker pool, kagent-default, with its one worker. That is the capacity my agents will run on. Below it come the actor templates and the actors. There are none yet, because there is no agent. That is the installation: kagent and Agent Substrate are running, and the cluster is ready for your first agent. In this lesson you learned to create a Kubernetes 1.37 cluster with kind, to install kagent from a config file and a values file, and to find kagent, Agent Substrate and the worker pool in the cluster.