What they did
Pierre-Laurent Medori runs backend engineering at GoodBarber, whose app builder writes app sections from a description. In October 2026 he worked with Claude Code to put a new engine, an agent that writes a section and tests it in a browser, into production for five test apps only. He set a rule that nothing could affect apps outside the allowlist. Claude Code wrote the gate functions, kept verbatim copies of old code paths for unlisted apps, wrote a management command to open the on/off levers in one transaction, and a script that proves each copy matches the code running in production. It also flagged what an allowlist cannot hold back: migrations, a scheduled job on the shared worker, and a proxy column. On the switch-on day he told it to push and deploy the config changes without him, step by step, with a check command before and after each step.
How it went
Fourteen of eighteen settings were on by 14:25 after four builds, with no failing checks; three stayed blocked and one was held back by choice. The proof script reported all 70 legacy copies identical to production. There were slips. Claude's own handoff note told him to open levers that did not exist yet, so for a day his tests silently ran on the old engine. And when he deployed a migration against Claude's advice, the build also shipped an unreviewed task from a colleague's session.
Worth knowing
Read how each gate treats an empty list: in one file, four gates read empty as everyone and three as no one. Claude made the agent's gate default to no one on purpose. Because the image builds from the tip of master, every switch-on cost a full build, so master had to stay frozen during the rollout. This is the engineer's own write-up; there is no outside account.