Kubernetes choices go beyond kagent: ARK and Orloj define Agent CRDs; Orka hosts coding runtimes. kagent 1.x splits agent definition (AgentTemplate) from execution policy (Harness), unlike its 0.x Agent YAML. Oracle Agent Spec is portable but needs a runtime adapter.

Kubernetes resources and portable specifications

kagent: check the versioned API

kagent 0.x has an Agent CRD (kagent.dev/v1alpha2) and sandbox variants. kagent 1.x changes this model: an AgentTemplate declares model, prompt and tools; a Harness declares the runtime and infrastructure policy; an AgentInstance is a running conversation created from the pairing. The Harness can run kagent, Codex, Claude or a bring-your-own A2A image, backed by Agent Substrate. This is a ready-made Kubernetes language, but 1.x is not simply one self-contained YAML document per agent. Verify cluster version and Substrate prerequisites before assuming deployment on the existing k3s. kagent 1.x core concepts | kagent 0.x API

ARK

The ARK Agent custom resource (ark.mckinsey.com/v1alpha1) carries a prompt, model reference and typed tool references (including MCP, HTTP or other agents). The runtime executes through a Query resource. A Query is an invocation record, not necessarily the team's canonical task board; a chat adapter could call the agent and return results in the original thread. The tool reference is not by itself a guarantee about the target system's credentials and authorization. ARK Agent reference

Orloj

The orloj.dev/v1 Agent has prompt, model_ref, available tools, roles and limits. Orloj also defines AgentRole, ToolPermission and AgentPolicy resources, with runtime governance. Be precise about allowed_tools: tools listed there are pre-authorized and bypass the AgentRole/ToolPermission check, although AgentPolicy restrictions still apply. A strict separation of who can declare an agent and who can grant tool use therefore needs a deliberate policy. It can be run with an operator or through the Orloj server, subject to version and deployment checks. Agent resource | Governance

Other Kubernetes-native formats

Orka Agent targets coding runtimes such as Codex and Claude through an Agent/Task model; assess its experimental maturity and whether Task is only execution state in the proposed integration. Sympozium resources include channels and policy, but role prompt may reside outside its Agent CRD. Inference Gateway ADL is a YAML language that generates an application and manifests; do not confuse its source file with the operator's separate CRD schema.

Portable and framework formats

Oracle Open Agent Specification defines a framework-neutral JSON/YAML representation of agents and workflows, with adapters. It is closer to a portable description than a controller for k3s; infrastructure permissions remain the deployer's work. Google ADK Agent Config describes agents in YAML but its declarative mode is experimental. Microsoft Agent Framework declarative agents are YAML/JSON loaded by its runtime. CrewAI, AutoGen and Haystack have native serialized configurations with differing code and workflow dependencies.

Selection test: compare the native schema and enforcement, local run path, chat invocation, k3s prerequisites, and whether execution resources merely record runs or try to own the human task lifecycle. The chat or optional GitHub Issue should remain authoritative for human work status.