kagent-dev/kagent

kagent – Kubernetes‑native framework for AI agents

What it is – kagent is an open‑source project that lets you define, run and manage AI agents as native Kubernetes resources. An agent (think of a chatbot or autonomous tool‑using service) is described in a YAML custom‑resource, and kagent takes care of launching the appropriate runtime, wiring it to LLM providers, and persisting its conversation state.

Why it matters – Running LLM‑backed services at scale usually means dealing with containers, networking, scaling, and observability. kagent folds all of that into the Kubernetes control plane, so you can use the same kubectl/GitOps workflow you already use for other cloud‑native workloads.

Core concepts (from the README)

Concept	Role
Agent	A Kubernetes custom resource that bundles the agent’s behavior (prompt, tools, sub‑agents) and runtime config.
AgentTemplate	Re‑usable definition of system prompt, tools and LLM settings that multiple agents can share.
Harness	Chooses the execution environment – Go ADK, Python ADK, Codex, Claude, or a custom runtime.
LLM Providers	Built‑in support for OpenAI, Azure OpenAI, Anthropic, Google Vertex AI, Ollama, plus custom gateway models.
MCP Tools	A catalog of “Managed Control Plane” tools (Kubernetes, Istio, Helm, Argo, Prometheus, etc.) that agents can invoke via a RemoteMCPServer resource.
Sessions	Persistent conversation/history stored in PostgreSQL; you can pause, resume, checkpoint or fork a session.
Observability	OpenTelemetry tracing of agent, model and tool calls, integrable with existing monitoring stacks.

Architecture at a glance

UI  →  kagent Controller  ←  CLI
          │                │
          ▼                ▼
   Kubernetes resources   gRPC / A2A / MCP APIs
          │
          ▼
        DB (PostgreSQL)   ← stores sessions & task history
          │
          ▼
      Substrate          ← lifecycle manager (suspend, resume, snapshots)
          │
          ▼
   Agent Runtimes (Go, Python, Codex, Claude, …)

The controller watches the custom resources, prepares agents, and talks to the substrate which actually runs the chosen runtime. A web UI and a CLI give users interactive access.

What you can do with kagent

Declare an agent in YAML, version‑control it, and let Kubernetes spin it up.
Swap models or runtimes by changing the Harness or ModelConfig – no code change needed.
Expose tools (e.g., Helm charts, Prometheus queries) to the agent via the MCP server, enabling autonomous operations.
Persist and replay conversations, useful for debugging or for building “checkpoint‑and‑continue” workflows.
Observe every step with OpenTelemetry, feeding data into Grafana, Jaeger, etc.

Getting started & community

Docs: Your first agent and Installation guide on https://kagent.dev.
Interactive UI demo (GIF in README) shows chat and tool management.
Contribute via GitHub issues, PRs, Discord, CNCF Slack channel, or the project’s Code of Conduct.
The project is CNCF‑incubated, Apache‑2.0 licensed, and follows OpenSSF Best Practices.

Bottom line – kagent is a genuine, production‑grade framework that brings AI agents into the Kubernetes ecosystem, giving developers a declarative, observable, and extensible way to run LLM‑powered services at cloud‑native scale.