AI Agents on Kubernetes Clusters

Kubernetes already provides what agents need to run safely at scale: declarative resources kept in Git, role-based access control, namespaces, service identity and network policy, autoscaling and a mature observability stack. The kagent project treats agents as ordinary cluster workloads that teams deploy, observe and govern with kubectl, GitOps, RBAC and a service mesh. There are two related uses: running agents on a cluster, and using agents to operate the cluster itself.

kagent is an open-source runtime for AI agents on Kubernetes, created at Solo.io and now a CNCF Sandbox project under the Apache 2.0 licence. According to its site, a control plane manages Agent and Session custom resources, agents are declared and versioned in Git, tools are exposed through an MCP layer, and agents can delegate to each other over A2A. It offers Go and Python agent runtimes and works with frameworks such as LangGraph, CrewAI and Google ADK. It emits OpenTelemetry traces and Prometheus metrics, can use Istio or Ambient Mesh for mTLS and egress policy, stores state in Postgres, and supports model providers including OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, Vertex AI, Ollama and Hugging Face. The quick install downloads a script from the repository and pipes it to bash, so read it first.

Agents and assistants that diagnose problems, explain errors and, with approval, change cluster state.

MCP servers let any MCP-capable agent query and manage clusters. Prefer read-only modes and scoped credentials.

Controlling how agents reach tools, models and each other, and isolating what they run.

Running the models the agents call on your own GPUs.

Treat agents like any workload that can cause harm: trace what they did, limit what they can touch, and require approval for risky actions.

Neighbouring pages on this site.