The Era of AI-Driven Development

WantedLab Tech Blog shared a case study on updating the iOS design system Montage to version 4.0 using an AI agent-based pipeline. Montage is an open-source project managed as a SwiftUI package, featuring approximately 60 components, 62 documentation pages, and 146 releases.

The core concept is an agent loop that connects designer tickets from implementation through Draft PRs, TestFlight deployment, and QA requests, and converts issues found during QA into new tickets to re-enter the same workflow.

• Create branches and implement based on tickets from the issue tracker and Figma specs
• Regenerate API documentation, commit changes, and create Draft PRs
• Deploy the component showcase app to TestFlight
• Share change logs and request QA via team channels
• Convert QA issues into new tickets to re-enter the loop

Developers do not implement directly within the loop; instead, they review and approve PRs or instruct on modifications before handling the merge. Releases are operated separately from the iterative loop, with version release events triggering the deployment workflow for the public documentation site.

With this approach, the median time from ticket creation to merge decreased from 34.9 days to 3.9 days. However, analysis suggests that most of the improvement stemmed from the elimination of wait times between tasks rather than an increase in implementation speed itself.