Project Lightning Talk: What's New And Exciting In Kagent? - Peter Jausovec, Maintainer

This talk introduces K agent, an open-source framework designed for running AI agents on Kubernetes. The speaker, Peter, explains that K agent, which has been part of the Cloud Native Computing Foundation since last year, includes a Go controller, a database, a UI, and a CLI. Developers can create agents using a declarative approach, allowing them to simply write instructions without writing code, or opt for a BYO option to build custom agents. K agent supports multiple AI models, including OpenAI, Anthropic, and AWS Bedrock, as well as tools for model management. Notable features added in the past year include human-in-the-loop capabilities, long-term memory storage, and support for both Python and Go runtime. Attendees are encouraged to try K agent and contribute to its development.

Hi, my name is Peter. I'm from solo and one of the maintainers of K agent. So K agent is an incubating project. We uh are part of the CNCF since last year. Since then we had 100 20 plus releases, a bunch of stars, and I'll tell you first what K agent is. So K agent is an open source framework for basically running AI agents in Kubernetes. So

very simple. Uh here's a more detailed look into um the infrastructure uh sorry of the architecture of the project itself. So it comprises of a go controller that's installed into your uh cluster. Uh there's a database, there's a UI portion, and there's a CLI. Um the nice thing about it is we also configure a couple of CRDs for you to use. So if you want to create

an agent and deploy it in Kubernetes, you can you have two options. You can use a declarative agent, meaning you don't have to write any code. You just write your instructions. You reference specific tools from MCP servers that you want to use. Uh you can reference skills, and then you you basically pick a model that you want your uh agent to use. We support multiple model models,

everything from OpenAI, Anthropic, Azure, AWS Bedrock, Llama, as well as AI Gateways. So if you're running any of the AI Gateways that front your models, you can point your K agent's agent through that as well. 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, and then build

your agents as docker images, and we can run that as Uh so here's how the CRDs uh look like. Uh so left side agent uh is the agent CRD uh where you can pick a model that you want to use uh that you deployed earlier on your cluster. You define your system message. We also support uh uh prompt templates, so you can define your templates and then

reuse them between different agents. And then after that we have a tools definition where you basically pick and choose which tools you want your agents to use. On the right-hand side, we have an MCP server uh definition where you can uh either point to an existing MCP server that's already running somewhere. In this case, we have the one that's deployed in the K agent namespace. And then

as you deploy this resource, it will automatically discover the tools for you uh from the MCP server. And then lastly, the model config, which is basically you're telling us what the provider is, what's the model, and you point to uh secret in Kubernetes that stores your API key. All right, so 1 year by the numbers. Uh as I said, March 1st March 4th was the first release.

We had more than 130 releases since, 204,000 2,400 stars. We have more than 100 contributors as well, a lot of forks, and currently we I think today we pushed a 0.8 version uh of the of the project. We have weekly meetings, bi-weekly contributor meetings, and we have a Discord that's very active. Uh so what what what what did we add in this past year? Here's a couple

of like high-level features. First one was hu- human in the loop, not humane in the loop. Uh which basically allows you to either confirm or do like a free-form input mode when your agent needs any input. Long-term memory storage, built-in prompts and templates that I showed. Uh we have both Python and go runtime selection. So for your declarative agents, you can either pick go or Python. We

also do get-based skill fetching. So if you have your skills that are pushed into the GitHub repos, we can do that in addition to OCI images. We do distributed tracing, A2A trace propagation, voice support, and bunch of other things. These are just some of the top-level ones that I wanted to call out. Uh join us so you can try the K agent at kagent.dev or if you

scan the QR code. And we would love to have you as a contributor. We are doing a talk tomorrow at the ContribFest. It's in G107 at 2:30 p.m. Thank you very much.