kagent-dev/kagent

Overview

kagent is a Kubernetes-native framework for building and running AI agents. Agents are modeled as Kubernetes custom resources, so their prompts, tools, model configuration, and runtime are declared in YAML and managed with familiar kubectl and GitOps workflows. It targets platform and DevOps teams that already operate Kubernetes and want agents to live alongside other workloads rather than in a separate silo.

The framework separates several concerns: Agents hold behavior and runtime config, AgentTemplates share prompts and tool sets across agents, Harnesses choose the runtime (Go or Python ADK, Codex, Claude, or custom), and ModelConfig resources select LLM providers. Agents can call MCP tools, including a built-in server covering Kubernetes, Istio, Helm, Argo, Prometheus, Grafana, and Cilium. Sessions persist conversation history and support suspend/resume and checkpoint-for-k forking. OpenTelemetry tracing covers agent, model, and tool execution.

Getting started

The documented path is a two-step flow: follow the installation guide to set up kagent on a cluster, then work through Your first agent to create and run an agent. The README does not embed the install commands themselves, so the exact prerequisites, cluster version, and CLI steps should come from the installation page. A web UI and a CLI are both provided for creating agents and chatting with them, and the controller exposes gRPC, A2A, and MCP APIs for programmatic access.

Architecture and operating model

The controller watches kagent custom resources, prepares agents, and manages their sessions; a web UI and CLI sit in front of it. Runtime work is delegated to Substrate, which handles agent lifecycle including suspension, resumption, and snapshots. PostgreSQL stores sessions, tasks, and conversation history independently of the running agents. Agent runtimes are selected through the Harness abstraction, which is what makes the framework extensible to non-ADK runtimes.

Constraints

Because agents are Kubernetes resources, a working Kubernetes cluster is a baseline requirement, and PostgreSQL underpins session persistence. Model access requires configuring a provider via ModelConfig; supported providers include OpenAI, Azure OpenAI, Anthropic, Google Vertex AI, and Ollama, with availability depending on the chosen Harness. The project describes itself as under active development, so APIs and resource definitions may change; the roadmap is tracked on the project Kanban board. Licensing is Apache-2.0.

HotRepo's take

kagent appeared on the Go daily list on 2026-11-07 at 10th position with 8 stars gained that day, and 4,177 total stars at the time. That is a single snapshot rather than a trend, and it shows the project attracting attention in the Go ecosystem even though its core value proposition is Kubernetes operations rather than Go-specific tooling. The most defensible technical observation is the split between declarative agent resources and pluggable runtimes, which lets teams keep agent definitions in version control while swapping the execution harness.