Why Refactoring Now Beats Waiting for Better AI Agents

Denys Linkov's team refactored a legacy 10-repo AI pipeline into a monorepo, betting that acting now would unblock growth despite rapidly improving models. The decision proved correct: shipping velocity surged, developer engagement increased, and new features that once took months now ship in under a week.

The tradeoff is clear. Waiting might have made the refactor cheaper, but the speaker's experiments show that even modern models like GPT 5.5 fail at zero-shot legacy rewrites—producing scaffolding without real implementation. Full automation remains unreliable for complex codebases.

Crucially, the talk highlights a missing mental model: most teams calibrate agent expectations at 50% accuracy, but real productivity requires 90%+ confidence. Until models consistently hit that bar, human workflow design, guardrails, and architectural decisions like monorepos drive success more than model capability.

Engineering leaders should take away that AI tooling is a force multiplier, not a substitute for sound architecture and governance. The refactor's patterns—monorepo, clear specs, human PR reviews—made AI agents effective. Waiting for perfect AI would have delayed the organizational benefits of cleaner code and higher developer morale.