The Call Center Doctors: 16,885 agent changes in six months, 44% merged overnight
A small call center company publishes six months of logs from an agent fleet that never stops: 158 merges a day, a night shift that never goes quiet, two fresh AI reviewers per change at about a dollar each, and a review queue that blocks almost as often as it merges overnight.
Source: 16,885 changes in six months. How a fleet of AI agents builds our software.
The Call Center Doctors, a call center company with a small team, describes how an agent fleet builds its software, with numbers from its own logs. From April 22 to October 21, 16,885 reviewed changes merged into its main code base, an average of 158 a day from August on. Between August 1 and October 21, 44% of merges (5,412 of 12,191) landed between 7 PM and 7 AM Eastern. The busiest hour was 1 PM, with 775 merges over the period. The quietest, 6 AM, still had 237.
The setup puts people at the two ends. They write down what they want and how they will know it worked. In between, a planning agent hands out jobs to as many as 40 helper agents at once. A shared sign-up sheet stops two helpers from editing the same files. Each helper must write a test that fails on the old code and passes on the new. Since October 18, important changes go through a stricter review: two fresh AI reviewers, each of which runs the tests and writes down every attack it tried, and a merge robot that merges only when both said SHIP on the exact same version. Those reviews took a median of 4.5 minutes and cost a median of $1.07 each at list price. Before the reviewers read anything, the kit plants small deliberate bugs in a throwaway copy of the code. The tests caught 55% of 3,755 of them, and the misses tell the reviewers where to look first.
The volume did not come from bigger changes. The median merged change shrank from 305 lines in August to 125 in October. The median time from opened to merged fell from 95 minutes in April to 2.5 in August, then rose again as reviews got stricter, which the authors say was a deliberate trade.
Why it matters for code review
This is one of the more detailed public records of a pipeline that runs around the clock, and its night log shows what the reviewers are doing at that hour. On one night in October, the log recorded 209 merge lines, 203 block lines and 277 stop-or-refusal lines. The overnight work produces almost as many rejections as merges, and steady throughput depends on review that keeps saying no. Across October 6 to 22, 17% of review verdicts were BLOCK and 16% of jobs went back at least once. The authors are open about the gaps: merge totals cover the main code base only, the token counts are a lower bound, and the strict-review figures cover just five days.