KubeStellar: 81% PR Acceptance in 82 Days Using AI Coding Agents

KubeStellar’s CNCF Sandbox Kubernetes dashboard project ran two AI coding agents — Claude Code (Opus) and GitHub Copilot — for 82 days. The result: 81% PR acceptance rate, 91% test coverage across twelve shards, bug fixes merged in under 30 minutes around the clock. The central finding was not about model capability. It was about what surrounds the model.

Context

Platform and engineering teams evaluating AI coding agents tend to frame the question around model selection: which model produces the best code, scores highest on SWE-bench, handles the most complex refactors. The KubeStellar Console experiment reframes the question entirely. Its author, Andy Anderson, argues that “the intelligence of an AI-driven development system resides not in the AI model itself, but in the infrastructure of instructions, tests, metrics, and feedback loops that surround it.”

This matters for cloud platform teams because it shifts the locus of investment. The model is a commodity input. The codebase infrastructure — documentation, test coverage, CI/CD automation, feedback loops — is where leverage accumulates.

The Detail

Anderson developed a five-stage AI Codebase Maturity Model from the KubeStellar experience:

Stage 1 — Instructed: Externalize agent preferences through explicit documentation. This means CLAUDE.md files, GitHub Copilot instruction files, and other structured prompts that define conventions, boundaries, and context the agent needs to work effectively. The agent only knows what the codebase tells it.

Stage 2 — Measured: Establish tests as the trust layer, not just correctness verification. Anderson’s formulation is precise: “A flaky test in a human workflow is tolerable. In an autonomous one, it’s a slow erosion of trust.” The KubeStellar codebase reached 91% coverage across twelve test shards, with 32 nightly test suites and 63 CI/CD workflows — not as targets, but as preconditions for agent autonomy.

Stage 3 — Adaptive: Automate only after measurement capabilities exist. Automation without measurement produces confident drift.

Stage 4 — Self-Sustaining: Let the codebase become the operating manual. At this stage, agents can onboard new functionality from the codebase state alone, without extensive human instruction for each task.

Stage 5 — Questioning: Ask “why” rather than “what.” Agents at this level surface root-cause issues rather than just executing instructions, enabling the system to improve its own constraints.

The KubeStellar metrics map to what this infrastructure enables in practice: community bug reports moved to merged fixes in approximately 30 minutes; feature requests became pull requests in roughly an hour. These are not averages across slow and fast tasks — they are consistent operational patterns across 82 days of continuous agent activity.

KubeStellar Console is built on Go (backend), React/TypeScript (frontend), and Helm packaging — a standard cloud-native stack. The infrastructure that produced these results is not exotic; it is disciplined.

Source: The New Stack

Content created with AI assistance and reviewed for accuracy.