kagent

kagent is a robust, CNCF sandbox project designed for DevOps and platform engineers to integrate agentic AI directly into Kubernetes clusters. This skill provides comprehensive support for users operating the kagent lifecycle, including CLI management, YAML-based agent declarations, and custom ADK-based agent development. Whether you are orchestrating LLM providers like OpenAI, Anthropic, or Ollama, or connecting complex toolsets via the Model Context Protocol (MCP), this guide enables you to build scalable, observable, and highly available AI agent architectures.

• Full support for the kagent CLI including installation, kagent get, kagent invoke, and kagent dashboard commands for observability.

• Detailed guidance on defining agents as Kubernetes custom resources (CRDs), configuring ModelConfig, and managing RemoteMCPServer or MCPServer instances.

• Configuration instructions for IDE integration (Cursor, Claude Code, Windsurf) using the controller’s /mcp HTTP endpoint for seamless agent-to-IDE communication.

• Implementation of the A2A (Agent-to-Agent) protocol for agent discovery, task orchestration, and inter-agent communication within a cluster.

• Best practices for system prompt design, managing tool permissions, and utilizing OpenTelemetry for tracing and debugging agent execution flows.

• The skill focuses exclusively on the user experience—CLI commands, Helm charts, and Kubernetes manifests—and does not cover internal project compilation or build-time development (e.g., make, go build).

• Always verify specific CRD field names and flag syntax against your currently installed version, as kagent evolves rapidly; use kubectl explain or kagent <command> --help to confirm details before deployment.

• Prioritize using standard Kubernetes primitives (kubectl, YAML manifests) to manage agents to ensure persistence and GitOps compatibility.

• Always set the required environment variables (API keys and providers) securely within the cluster or environment when deploying custom models.

• When troubleshooting, rely on kagent bug-report and controller logs (kubectl logs -n kagent) to diagnose pod failures or MCP handshake issues.