Skip to content- Sources: primary
- Summary: Kush Agrawal describes Temporal workflows that score files from Xcode coverage reports, run one Claude Code agent per file to open a labeled PR, and fix red CI up to five times before giving up. The workflows assign the top git blame contributor as reviewer with auto-merge on. A PR that edits non-test files is closed automatically. An LLM judge has reviewed about 1,200 PRs and rejected 73 for edited production code, deleted or disabled tests, or long waits. Coverage of the app's MVVM layers rose from 9% to 59% and overall coverage from 11% to 24%. Lint failures dropped once the agents moved from Linux to Mac runners and could build locally.
- Why it matters: It shows a concrete identify, write, verify and resolve loop for unattended agent PRs, including the guardrails that stop agents from gaming CI.