Solo.io has made enterprise editions of Agent Substrate and kagent generally available, aiming to help organizations build and run AI agents on Kubernetes clusters while managing resource use and access controls. The company announced the availability at KubeCon + CloudNativeCon North America 2026 on Nov. 13.
Agent Substrate is a framework jointly developed by Solo.io and Google. It is designed to separate an agent’s stateful session and identity from the compute resources needed to run it. When an agent is idle, its state can be captured and stored; the session can then be restored in under 500 milliseconds when activity resumes, according to the announcement. The approach is intended to avoid keeping Kubernetes pods occupied by agents that are not actively working.
Solo.io’s enterprise edition gives teams a supported option for adopting the framework before it is incorporated into Kubernetes distributions. Agent Substrate has been donated to the Cloud Native Computing Foundation (CNCF), and the company says the donation should lead to its inclusion in distributions in 2027. The announcement does not specify which distributions will include it or when they will do so.
Managing agents as a Kubernetes workload
Agent activity can be difficult to predict. A large number of agents may become active at once, consuming cluster capacity and potentially occupying available pods. Separating stored session state from active compute is intended to make it possible to scale compute use more closely to actual activity, rather than reserving resources for every persistent agent session.
The framework’s runtime, called an actor, provides an isolated execution sandbox and a gateway that governs how the actor is reached and what it can access. State, identity and access controls remain associated with the actor, according to Solo.io. This design makes those controls part of each session’s execution environment rather than treating them as separate concerns.
Solo.io’s kagent is a tool for building, deploying and managing AI agents. The company says kagent can now invoke Agent Substrate, enabling teams to use the agent-development tool alongside the framework’s session and compute-management capabilities. The announcement positions the two generally available products as components for organizations seeking to run agents on infrastructure they already operate, instead of introducing a separate platform for that workload.
Identity and security controls
The broader set of capabilities described by Solo.io also includes agentgateway and agentshield. Agentgateway can enforce identity and policy controls on incoming and outgoing traffic. Agentshield is intended to detect attack paths and help protect the environment. Together with the actor’s sandbox and access controls, these tools are aimed at addressing governance and isolation needs as agent deployments grow.
Keith Babo, Solo.io’s chief product officer, said IT teams face pressure to deploy AI agents safely at scale. The company’s announcement focuses on runtime efficiency and security, but it also points to a wider infrastructure challenge: agents can generate substantial traffic to application programming interfaces (APIs). Kubernetes control planes may need changes to handle the volume of API calls associated with large enterprise deployments, the announcement notes.
That control-plane issue is distinct from suspending idle sessions. Agent Substrate addresses the relationship between an agent’s persistent state and active compute, while Kubernetes still has to coordinate the workloads and API activity across a cluster. The announcement does not detail specific control-plane changes or provide performance benchmarks for handling agent-generated API traffic.
For enterprises, the immediate question is how to accommodate agent workloads without letting unpredictable activity consume cluster capacity or weakening oversight. Solo.io’s availability announcement offers a supported implementation of Agent Substrate and a connected agent-building tool, while the CNCF donation sets a separate path toward possible inclusion in Kubernetes distributions in 2027. Adoption and integration will determine how the approach fits into existing cluster operations.