Kagent Review

Teams that run everything on Kubernetes don't want AI agents to be a separate, ungoverned island with its own deployment and security story. kagent's pitch is to end that: run agents natively on Kubernetes, the way you run everything else.

Agents become Kubernetes resources, with multi-LLM support, MCP tool servers, observability, human-in-the-loop gates and sandboxing, all installable with a single helm command. It sits among agent frameworks like Cline, but at the cluster-infrastructure layer.

This review covers how kagent runs an agent, which teams it suits, how it's licensed, and where its fit is specific, to weigh first.

kagent Kubernetes-Agent Facts

Running AI Agents Natively On Kubernetes

kagent's core idea is to make AI agents first-class citizens of Kubernetes. Agents are defined as Kubernetes resources (CRDs), so you deploy, version and manage them with the same GitOps, kubectl and RBAC you already use, rather than bolting on a separate agent platform.

It's model- and framework-agnostic: it works with many LLM providers (OpenAI, Anthropic, Gemini, xAI, Azure OpenAI, Bedrock, Vertex, Ollama, Hugging Face) and lets you bring frameworks like LangGraph, CrewAI or Google ADK, connect MCP tool servers, and use agent-to-agent (A2A) communication.

For operations it brings the things platform teams expect, OpenTelemetry and Prometheus observability, long-term vector memory, human-in-the-loop approval gates, context compaction, sandboxing and RBAC, Istio/Ambient mTLS, and Postgres storage, all installable with a single helm command.

The honest framing is that kagent is a serious, well-pedigreed piece of infrastructure, it's open source under Apache 2.0, a CNCF Sandbox project created at Solo.io (whose team founded Istio), with a growing GitHub following. Because it's open source and self-hosted, there's no lock-in and every feature works without waiting for an enterprise deal. The caveats are about fit, not quality: it assumes you already run Kubernetes and have the platform-engineering skills to operate it, so it's not for individuals or teams wanting a managed, no-ops agent service; and as self-hosted infrastructure, the running and securing of it is on you. Its claims are vendor-stated, though the code being open helps you verify them.

kagent is serious, open-source agent infrastructure for Kubernetes teams; its only real limit is that it assumes you run Kubernetes and can operate it.

How Does kagent Run An Agent?

kagent pairs Kubernetes-native agent definitions with model and framework freedom.

Agents As Kubernetes Resources

Agents are Kubernetes resources managed by GitOps, kubectl and RBAC, so they fit your existing cluster operations.

An agent coding tool like Relevance AI offers a managed, low-code agent platform; kagent's angle is self-hosted, Kubernetes-native infrastructure, the opposite end, for teams who want to own the stack.

Models, Tools And Observability

It's model- and framework-agnostic, with MCP tools, observability, memory and human-in-the-loop gates, installed via one helm command.

A coding agent like Kilo Code helps you write software; kagent is about running agents in production on your cluster, so match it to whether your need is building or operating agents.

Which Platform Teams kagent Suits

kagent suits platform-engineering, DevOps and SRE teams that already run Kubernetes and want to deploy and govern AI agents with the same tooling, controls and observability as their other workloads.

It suits organisations that value open source and self-hosting, no lock-in, full control, and that have the skills to operate cluster infrastructure and want model and framework freedom.

It suits less well individuals or small teams without Kubernetes, anyone wanting a managed, no-ops agent service, and non-technical users, this is infrastructure, not an app.

A fair way to place it: if you run Kubernetes and want agents to be a governed, observable, self-hosted part of your platform rather than a separate SaaS, kagent is an excellent, no-lock-in fit; if you don't run Kubernetes or want someone else to operate it, a managed agent platform will serve you far better, which is purely a question of fit.

kagent's Licence And Enterprise Tier

kagent is open source under Apache 2.0 and free to self-host, every feature works with a single helm install, with no paywall; an 'Enterprise' option exists but no price is published. We do not publish prices; the open-source core costs nothing beyond your own infrastructure and operating effort.

The value question isn't licence cost (the core is free) but operating cost: running agents on your own cluster means you bear the compute, the model-API usage and the engineering time to operate and secure it.

The sensible approach is to install the open-source version on a test cluster, wire up a model provider and an MCP tool, run a real agent through its observability and approval gates, and judge the operating effort before standardising on it.

Deploying One Agent With kagent

Because it's self-hosted infrastructure, deploy one agent on a test cluster first.

1. Install kagent on a test cluster with a single helm command.
2. Wire up a model provider and an MCP tool server.
3. Deploy one real agent as a Kubernetes resource via GitOps.
4. Exercise the observability, memory and human-in-the-loop gates.
5. Judge the operating and securing effort for your platform team.
6. Pass/fail: kagent is worth it if it fits your Kubernetes operations.
7. Treat capability claims as vendor-stated; the open code aids verification.

kagent, The Agent-Runtime Brief

• kagent runs agents on Kubernetes: AI agents as native Kubernetes resources.
• It's open source: Apache 2.0, a CNCF Sandbox project, self-hosted.
• It's built for platform teams: GitOps, RBAC, observability and sandboxing.
• The fit is specific: it assumes you already run Kubernetes.

kagent Kubernetes-Agent Terms

Kubernetes-native

Built to run as part of Kubernetes, using its resources, tooling and controls rather than a separate system.

CRD

Custom Resource Definition: a way to define your own object types (here, agents) that Kubernetes manages.

MCP

Model Context Protocol: a standard for connecting AI models to external tools and data sources.

Human-in-the-loop

A design where a person approves or checks the AI's actions at key steps rather than letting it run unchecked.