KAIF·k8s agentic ai fabric
Pillar 01 · build

Development Lifecycle

The development lifecycle is how agentic capability comes into existence. It covers authoring of agents, prompts, tools, skills, references and datasets; the harness configuration that wires them together; the automations that drive CI; and the registry where every artifact lands signed and versioned. A good loop is fast, reproducible, and ends with something the Deployment pillar can promote and the Evaluation pillar can score.

Modules involved
plan hub →
kaif-platformThe technology substrate every other module runs on.kaif-designBusiness use case → agentic design proposal, evaluated offline and online.kaif-evalOne evaluation spine, four surfaces — design, build, deploy, runtime.kaif-guardAll guardrails and policies. Publishes bundles kaif-deploy enforces.kaif-identityAuthN and AuthZ across users, agents, tools and tenants.
What good looks like

Outcomes

  • A new agent goes from idea to running-in-cluster in under a day.
  • Prompts, tools, skills, references and eval datasets are versioned alongside code in Git.
  • Every artifact — agent image, model, prompt bundle, skill pack — is signed and SBOM-attested.
  • Developers use the same Kubernetes primitives in dev and prod — no environment drift.
  • The Harness registry is the single source of truth for what exists and who owns it.
The loop

How it actually runs

  1. step 01
    Scaffold (Backstage)
  2. step 02
    Author in Git (code · prompts · skills)
  3. step 03
    Inner loop in cluster (Tilt)
  4. step 04
    CI build · sign · SBOM (Tekton)
  5. step 05
    Publish OCI artifact (Harbor)
  6. step 06
    Register in Harness
Cloud-native enablers

Technology on Kubernetes

Every capability maps to a CNCF or Kubernetes-native project. No parallel stack — the same cluster runs your agents and your platform.

Authoring · agents

  • LangGraphStateful graph-based agents as containers.
  • CrewAIMulti-agent crews with role specialization.
  • AutoGenConversational multi-agent orchestration.
  • Custom (Python / TS)Plain workloads when frameworks are too heavy.
  • Agent CRDDeclare an agent as a Kubernetes object.

Authoring · prompts, skills & references

  • Prompt bundles (OCI)Versioned prompt templates with variables and tests.
  • Skill packsReusable capability units (write-PR, run-tests, query-Jira).
  • Reference docs / RAG corporaCurated knowledge bases registered with provenance.
  • Promptfile / .promptFile-based prompts reviewed as code in PRs.
  • MCP serversStandardized tool / context endpoints the agent consumes.

Authoring · tools & datasets

  • Tool manifests (JSON Schema)Typed tool contracts shared across agents.
  • OpenAPI · gRPCExpose existing services as agent tools.
  • DVC / LakeFSVersion eval and fine-tune datasets.
  • Argo Dataflow / Spark OperatorBuild and refresh datasets on the cluster.
  • Vector ingestion jobsEmbed and load references into Weaviate / Qdrant.

Inner-loop tooling

  • Dev ContainersReproducible dev environments per agent.
  • Tilt / SkaffoldLive reload of agent pods on file save.
  • TelepresenceRun one agent locally against the cluster.
  • kind / k3dLocal Kubernetes that mirrors prod.
  • Devfile / DevWorkspacePortable cloud IDE specs.

Developer portal & automations (CI)

  • BackstageSelf-service templates and a catalog of agents, skills and tools.
  • TektonKubernetes-native CI pipelines.
  • Argo EventsTrigger builds and evals from Git, chat, webhooks.
  • Argo WorkflowsDAG automations for build · test · eval · publish.
  • DaggerPortable pipeline code that runs locally and in CI.
  • RenovateAutomated dependency and model-version bumps.

Artifacts, registry & supply chain

  • HarborOCI registry for images, charts and ML artifacts.
  • KitOps / ModelKitsPackage models + prompts + datasets as OCI artifacts.
  • Sigstore / cosignSign every artifact and verify on admission.
  • in-toto / SLSAProvenance attestations for the full build chain.
  • Syft / TrivySBOM generation and vulnerability scanning.
  • External Secrets + VaultInject API keys without leaking them.
Cross-cut

How the Harness plugs in

The Harness is itself authored here. Its catalog services — Agent Registry, Prompt Catalog, Skill Library, Reference Index, Tool Directory and Dataset Index — are declared as Kubernetes CRDs and updated by CI. When automations publish a new version, it lands in the registry with metadata, signatures and ownership, and becomes immediately available to the Deployment pillar via GitOps and to the Evaluation pillar as a target.
Deployment Lifecycle →