What they did
Nestor Cortina migrated two production apps with coding agents: an Ionic mobile app to React Native plus a new TypeScript facade server in front of several old backends, and an AngularJS web portal to React. Before launching anything he wrote an AGENTS.md style guide with good and bad examples, per-area docs, task skills, a pre-commit hook that rejects files over 20 KB and runs lint and typecheck, and CI that fails on any warning. PRs were capped at about 500 lines and 15 files. He built the first module with a single agent and iterated on it until it was the reference port, then gave the other agents a one-page brief pointing at it. Two or three agents ran in parallel on the mobile app and nine on the web portal. Afterwards agents drove the app in an Android emulator and three cheap subagents walked all 84 web routes in the old and new portal to compare where each one landed. He does not name the agent product.
How it went
The agents wrote most of the code in a couple of hours. Review took a weekend and more: over 120 stacked PRs on mobile and one 54,000-line PR on the web, checked with read-only auditor agents. The server came to about 30 modules, 22,000 lines plus 9,000 lines of tests. The web login went from 2,677 to 852 lines after a style pass. Ten agents at once on the mobile repo produced no commits at all. Four web agents copied an earlier bad migration attempt and one module did not compile, so that work was thrown away and rerun with another model. Emulator testing found a navigation reset and an Android crash; the route comparison found a settings screen left unprotected. It still needs polish and an internal beta.
Worth knowing
Agents copy whatever patterns they find in the repo, including old or broken code, so a reference module and written rules mattered more than the prompt. How many agents can run at once depended on how much shared code the modules touched. An automatic conflict resolver saved time but sometimes broke JSON or revived deleted code.