30 to 70 PRs a Day: How We Managed to Not Wreck Our Systems
Honeycomb's peak weekday merges went from ~30 to ~74 in a year and incidents went 1.7× then 2.9×. What continuous delivery, flags, and ownership absorbed — and what they didn't.
Honeycomb's numbers, all from its own repo and incident log: at least 82.6% of new lines attributed to AI by June 2026, a codebase from 0.97M to 2.1M lines in sixteen and a half months, and a 2024 incident baseline of 18.5 a quarter against 53 in Q2 2026. Fong-Jones reads the incident curve as linear in change volume and notes that no AI-authored change caused a catastrophic outage. Adoption came in three steps tied to model releases, with Opus 4.6 in February 2026 the visible one, and to leadership pushing AI-first. The absorbers were continuous delivery, fast CI, observability that links code to production behaviour, and blameless review, all in place beforehand. Her caveats: peak weekday counts, attribution as a floor, and concurrent org changes she cannot separate out. The bill, by her account, is the incident count, and she paid it.
Comments
Loading comments…