Worth the Squeeze
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A Gusto engineer shares how agentic AI workflows made a large-scale Ruby codebase refactor feasible by replacing OpenStruct with faster alternatives. The key insight wasn't AI's code generation speed but the ability to cheaply iterate on the decomposition strategy itself. Initial attempts using one PR per module caused reviewer fatigue; grouping PRs by team ownership solved the bottleneck. Claude Code was used to build a reusable skill that scans for deprecated patterns, groups changes by team, generates replacements, and measures performance. A two-agent setup ran refactoring and performance benchmarking in parallel using git worktrees, producing concrete metrics (25–39x faster instantiation, 90%+ memory reduction) that justified the work. A GitHub issue served as a living dashboard and shared memory between sessions.
Table of contents
The Cost of Standing StillThe Naive Approach (and Why It Fails)Iteration 1: Group by PackIteration 2: Group by TeamIteration 3: Creating a Repeatable SkillGet Jose Miguel Colella ’s stories in your inboxThe Living DashboardScaling Up: Agent Teams and WorktreesWhat I LearnedWorth the SqueezeSort: