A K8s cluster build-and-recovery platform: an internal tool I proposed and built with AI agents

Building Kubernetes clusters from scratch and repeated lab work exposed their complexity and learning curve. To help IT staff provision and configure clusters, I began developing a Go management platform with AI agents in August 2025. A VM-based on-premises cluster on Proxmox takes an estimated 20–30 minutes, depending on storage I/O; backup and recovery capabilities are verified separately.

Track
Self-proposed
Status
POC passed
Period
2025-08 – 2026-08
AI
Developed with AI agents

My scope

Requirements, architecture and component choices, Kubernetes domain knowledge, acceptance criteria, and testing on physical hosts. The Go code was written by AI agents; I read, review and verify it.

Approach

  • Built with Claude Code; Go, React, OpenTofu, Proxmox API.
  • A single controller integrates the versioned API, SQLite / WAL workflows, OpenTofu / Proxmox provisioning and mTLS node agents. Cluster components and recovery capabilities are described by version and acceptance evidence.
  • Repeated build, node-join and restore tests on my own physical Proxmox hosts.

Verification

Full build and restore drills on physical hosts; a successful POC on the product team’s test cluster; all unit tests pass.

Results

MetricValueSourceEvidence
Build an on-prem VM-based cluster20–30 min (conservative)Own measurementsOwn estimate
Unit tests5,099 run: 5,020 passed, 79 skipped, 0 failedgo test -short, 2026-10-21System record
Statement coverage52.2% (domain layer 78.5%)Same runSystem record
Related hours656.5 hTimesheets, 2024–2026System record
Logged hours split: me / AI executionme ≈277 h, AI execution ≈380 hOwn estimate; hours since 2025-02-22 split by the period’s ratio; company-logged hours onlyOwn estimate

Limits and next steps

I rely on AI agents at the Go language level. Packages at 0% coverage are mostly entry points, generated gRPC code and test helpers. It ended up as a fast build-and-recover tool for test environments, not production.

Tools

  • Go
  • React
  • OpenTofu
  • Proxmox
  • RKE2
  • Longhorn
  • Calico
  • Claude Code