kagent_skill

Overview

This skill implements kagent, a Kubernetes-native AI agent framework for building, deploying, and managing AI agents as Kubernetes resources. It enables declarative agent definitions via Custom Resource Definitions (CRDs) so agents behave like first-class K8s objects. Use it to run persistent, cluster-aware agents that analyze, remediate, and automate complex Kubernetes operations.

How this skill works

Kagent installs CRDs and a controller that creates AI agent pods based on Agent, Tool, and Model CRDs. Agents can execute kubectl-style operations, use diagnostics (crictl, kubelet logs, tcpdump), and report findings back into Kubernetes. The framework ties model providers, tool permissions, and namespaces together so you can control scope, auditing, and lifecycle through standard Kubernetes workflows.

When to use it

• Deploy persistent AI agents that must be managed declaratively in Kubernetes.
• Perform automated cluster-wide health checks, resource optimization, or security audits.
• Automate complex operational runbooks and AIOps tasks with policies and staged rollouts.
• Integrate AI-driven observability and recommendations into CI/CD and incident response.
• Limit agent scope and permissions via CRDs when you need auditable, namespace-scoped automation.

Best practices

• Start with read-only agents and validate outputs before granting mutating permissions.
• Scope agents to specific namespaces and tools to least-privilege permissions.
• Enable logging and an audit trail for every agent action to track changes and approvals.
• Set CPU/memory limits for agent pods to avoid resource contention in the cluster.
• Adopt a gradual automation path: verify recommendations, then enable automated remediations.

Example use cases

• Cluster health analysis before large deployments: run agents to surface node, pod, and component issues.
• Resource optimization: detect pods without limits, recommend adjustments, and propose scaling plans.
• Security audits: find pods running as root, exposed services, or secrets mounted insecurely.
• Troubleshooting: trace root causes for OOMKilled pods, network failures, or failing deployments.
• Operational automation: implement a controlled runbook that cordons, drains, and remediates unhealthy nodes.