How Agents Rewrote a Node.js Service in Go (With a 600-Test Harness)

Checkly

Sep 28, 2026

Stefan sits down with Eddie, senior engineer at Checkly, to talk through how an AI agent rewrote a legacy Node.js service in Go, and what it actually took to ship that to production safely.

Instead of prompting an agent to "rewrite this, no mistakes," the team started with zero service test coverage and built a black-box end-to-end harness first (~600 tests) using Playwright and Docker Compose. A single Claude Code instance was then handed the harness as acceptance criteria and left running overnight, producing a 13,000-line Go PR. The team reviewed it, deployed to Kubernetes, and rolled out to customers in cohorts behind feature flags with instant rollback.

The result: 20 pods down to 5, roughly 15 vCPUs and 45 GB of memory saved, database transactions resolving 10–20x faster, and a migration that took about four weeks instead of the five to six months a similar rewrite took the year before.

Checkly: https://www.checklyhq.com
Checkly Docs: https://www.checklyhq.com/docs
Try Checkly for free: https://app.checklyhq.com

00:00 AI Rewrite Hype

00:50 What Checkly Does

02:34 Scaling Pain Points

05:09 Harness First Strategy

06:56 Harness Explained

10:47 Data Driven Test Cases

18:51 Testing Outages Toxiproxy

21:28 Agent Rewrite To Go

23:43 Reviewing The 13K PR

25:07 Deploy And Iterate

29:32 Harness Makes Determinism