The Kubernetes Production Gap for Agentic Workloads

▶ Watch (0:03)

AI developers want to run everything, everywhere. That works for pilots and POCs. Production is different. Kubernetes, unlike other agentic runtimes on the market, has no direct representation of agents, skills, or tools as platform entities. Solo.io built three open-source projects to fill this gap. K agent targets runtime identity, Agent Gateway handles networking and security, and Agent Registry provides catalog visibility across every environment.

K Agent: Agents as First-Class Kubernetes Entities

▶ Watch (0:42)

K agent treats agents, skills, and tools as first-class Kubernetes entities. It acts as a general-purpose agent harness with declarative agent definitions, similar to Open Claw. It also supports deploying agents built in LangChain, ADK, and CrewAI. On top of K agent, Agent Gateway adds the production controls developers need: memory management, session optimization, and human-in-the-loop controls. It also removes the undifferentiated infrastructure work platform teams would otherwise carry.

Agent Gateway: Four Requirements for Production-Grade Deployment

▶ Watch (1:25)

Solo built Agent Gateway around four requirements. First, it needs every feature teams need today, plus room for what the field will need next. Second, it must run at scale without consuming excess infrastructure resources. Solo published a public benchmark so teams can test this themselves. Third, Agent Gateway sits under the Linux Foundation, with multiple companies already contributing. Fourth, it needs real users running production workloads. Solo claims all four.

Agent Registry: Catalog Visibility and Live CNCF Donation

▶ Watch (2:44)

Visibility into deployed agents is a widespread problem. Agent Registry gives teams a single catalog covering agents, tools, skills, prompts, and evaluators across all environments. It supports federation from multiple private and public registries, with a curation workflow that ensures only approved artifacts reach developers. It exposes as an MCP tool server, so frameworks like K agent and Claude Code can query it with semantic search. Registered runtimes include K agent, Vertex, Agent Core, AI Foundry, and Nemo Cloud from Nvidia. On stage at KubeCon EU 2026, Solo donated Agent Registry to the CNCF.

Notable Quotes

Second of all, it need to be efficient. Idit Levine · ▶ 1:53

let’s donate Agent Registry to the CNCF, Idit Levine · ▶ 5:13

Key Takeaways

  • Kubernetes has no native agent representation; K agent fills that gap with first-class agent, skill, and tool objects.
  • Agent Gateway, in the Linux Foundation, ships with a public benchmark and handles memory, session, and human-in-the-loop controls.
  • Solo donated Agent Registry to the CNCF live on stage, providing a drift-detecting catalog across all agentic runtimes.