arkonis-operator

arkonis-operator is a Kubernetes operator that introduces AI agents as first-class cluster resources. Instead of managing containers, you manage LLM-based agents: each defined by a model, a system prompt, and a set of MCP servers that provide tools. The operator handles scheduling, scaling, health checking, and lifecycle management using the same reconciliation model as any other Kubernetes controller.

Everything integrates with the Kubernetes ecosystem you already use. GitOps workflows, RBAC, namespaces, and kubectl all work without modification. Promoting a new system prompt is a pull request; rolling it back is a revert in ArgoCD or Flux, the same workflow you already use for everything else.

CRDs

Resource Analogy Description
ArkonisDeployment Deployment Manages a pool of agent instances running the same model, prompt, and MCP tool servers. Handles scaling and semantic health checks.
ArkonisService Service Routes incoming tasks to available agent instances using configurable load balancing strategies (round-robin, least-busy, random).
ArkonisConfig ConfigMap Reusable prompt fragments and model settings. Referenced from ArkonisDeployment via spec.configRef to keep configuration DRY.
ArkonisPipeline (novel) Defines a DAG of agents where outputs feed into inputs. The primitive for declarative multi-agent workflows.
ArkonisMemory PersistentVolumeClaim Configures a persistent memory backend (Redis or vector store) for agent instances. Referenced from ArkonisDeployment via spec.memoryRef.

Short names: aodep, aosvc, aocfg, aopipe, aomem

Next steps


Apache 2.0 · ARKONIS