kagent is a Kubernetes-native runtime for AI agents, implemented as a CNCF Sandbox project under Apache 2.0. It installs as a controller in an existing cluster and exposes agent identity, policy enforcement, and lifecycle management as Kubernetes custom resource definitions.
Agents from any framework (LangChain, CrewAI, Google ADK, A2A protocol, or custom Python) are deployed as kagent workloads. The runtime assigns each agent a verifiable identity with on-behalf-of semantics, applies policy, and traces all actions and tool calls via OpenTelemetry. Human-in-the-loop approval workflows gate critical agent actions before execution.
MCP tool servers are deployed as sandboxed Kubernetes resources via the kmcp sub-project. agentgateway handles protocol routing between A2A and MCP; agentregistry provides skill discovery. Enterprise support, SLAs, and a hardened distribution are available through Solo.io. Requires existing Kubernetes expertise to operate.
