Faster AI-assisted coding shifts the bottleneck to CI verification and costs

A team with about 25 engineering agents found that writing code was no longer the main constraint, as verification through test runs became the bottleneck. Their GitHub Actions spending in October reached roughly $1,600, more than two and a half times what they paid for three Claude Code accounts. Repeated full test-suite runs on pull requests and merges drove most of the runner-minute billing. The article describes how to audit job runs via the GitHub API to find duplicated test executions. The team cut billed minutes per run by 59% by splitting testing into draft and merge-ready tiers, and by using a merge queue to avoid redundant full runs.