Kagent: Running AI Agents as Kubernetes-Native Resources

▶ Watch (0:02)

Kagent is an open source framework for running AI agents inside Kubernetes clusters. It ships a Go controller, a database, a UI, and a CLI. Teams also get a set of CRDs that make agents first-class Kubernetes objects. Two deployment paths exist: a declarative option that needs no code, and a bring-your-own-agent option built on frameworks like ADK, LangGraph, or CrewAI packaged as Docker images. The project joined the CNCF as an incubating project in 2024.

Declarative Agents: Models, Tools, and MCP Servers via CRD

▶ Watch (1:44)

The Agent CRD takes a model reference, a system message, and a tools list. Prompt templates can be defined once and reused across agents. A separate MCPServer CRD points at any running MCP server. Once deployed, kagent auto-discovers the tools that server exposes. A ModelConfig CRD stores the provider name, model name, and a Kubernetes secret reference for the API key. Supported providers include OpenAI, Anthropic, Azure, AWS Bedrock, Llama, and any AI Gateway sitting in front of those models.

One Year by the Numbers

▶ Watch (2:49)

The first release shipped on March 4th. Since then the project has produced more than 130 releases and collected 2,400 GitHub stars. Over 100 contributors have merged code, and version 0.8 pushed on the day of this talk. The team runs weekly community meetings, bi-weekly contributor meetings, and maintains an active Discord. That pace across a single year establishes kagent as one of the faster-moving CNCF incubating projects by release cadence.

New Features in the 0.8 Release Cycle

▶ Watch (3:23)

Human-in-the-loop support lets an agent pause and accept either a confirmation click or free-form text input before continuing. Long-term memory storage persists context across sessions. Both Go and Python runtimes are available for declarative agents. Skills can now be fetched from GitHub repositories alongside OCI images. Distributed tracing and A2A trace propagation ship built-in, giving operators end-to-end visibility across agent calls. Voice support rounds out the 0.8 feature set highlighted at the talk.

Notable Quotes

K agent is an open source framework for basically running AI agents in Kubernetes. Peter Jausovec · ▶ 0:20

today we pushed a 0.8 version uh of the of the project Peter Jausovec · ▶ 3:06

if you want more uh control over your agents, we also have a BYO option where you can build your own agents using either ADK, LangGraph, or QAI uh CrewAI Peter Jausovec · ▶ 1:27

Key Takeaways

  • Kagent deploys AI agents on Kubernetes using CRDs, with no application code required for declarative agents.
  • In one year: 130-plus releases, 2,400 stars, 100-plus contributors, and version 0.8 shipped at KubeCon.
  • Built-in distributed tracing and A2A trace propagation give operators visibility across multi-agent calls.

About the Speaker(s)

Peter Jausovec is a software architect at Solo.io with over 18 years in software development. He maintains the kagent project and has worked on AI gateway and agent development at Solo. His work connects service mesh and API management with AI workloads.