Request for new kubernetes-sigs repo for Agentic Networking

[Lior Lieberman]

Hey Everyone,

Over the last month, we've had some initial discussions about adapting agentic networking, and networking governance for AI workloads. It’s been exciting to see all the interest and momentum in this area. We’ve found a great group of people that are interested in making sure that Kubernetes networking is well equipped to support agents. In practice, this means ensuring we’ve provided the right policies and guardrails to support communication from agents to tools, and agents to agents, focusing on new protocols like MCP and A2A. To keep working towards this goal, we think it would be useful to spin up a new kubernetes-sigs repo where we can iterate together on new APIs and a basic prototype implementation of those APIs. 

We have a doc with a lot more information on what this work would entail. We’d like to build a big tent here, so if we’ve missed anyone that might be interested in this effort, please reach out.

If there are no objections, we will follow up with requesting a repo in kubernetes/org in the next few days. 

Thanks!

[Antonio Ojea]

Thanks for driving this, I think this is an important topic, and I agree having an implementation, can help us to take better decisions and iterate faster.

We have learned that not being able to meet the end users demands without iteration has led to fragmentation because users couldn't test implementations and provide feedback, and when implemented elsewhere it never came back to Kubernetes or it was just too late. We need to build an ecosystem of small and scoped components that fit well together and this proposal seems to go in this direction.

Just a few comments:

- I suggest that before we create the repo, we have a clear proposal for the name, a list of initial project owners, and a clear governance model. For the name, we always prefer boring names: kube-network-policies, node-ipam-controller, gateway-api-inference-extension ... simple and descriptive like gateway-api-agentic-extension seems like a good direction, I really would like to avoid catch-all names as "agentic networking" if possible.

- As expressed in the doc, this is a new area with overlap, I'd also like to be sure that all stakeholders are aligned, if there are any concerns those should be raised now so we can address them.

 

- Another important topic is the alignment with the entire SIG, we do not want to have different networking models, as "if this is network agent traffic do this, if is Pod to Service do the other, and if I have network policies is implementation specific ..." , so please remember that the SIG Network charter requires subprojects to provide regular updates to the SIG, and API reviews, just to help we communicate accordingly and we flesh out these details

Happy to +1 once this questions get clarified