What they did
Romain P., who works a full-time contract alongside, wanted to see how far a messy codebase could go if it were only steered from a phone. Grok Bot acted as project lead and the only contact. It drove Cursor and Muse, installed on a 32-core Scaleway server, to plan, code and review tickets day and night, following working rules already kept in the repo. Grok Bot held the keys to GitHub, npm, Cursor, Muse, Scaleway and the author's mailbox. The human side was short messages asking for status, asking for a branch to be deployed to check on the phone, or sending a screenshot of a broken screen with a comment. The author wrote no code during the experiment.
How it went
Between 5 and 14 November 2026 the server logged 1,541 agent sessions, 988 of them code reviews, about 180 agent hours and roughly $1,390 of usage at list price. The project gained a repo, a site, docs, published npm packages and a public page showing each ticket being planned, built and reviewed. The misses: the project was public with no lab label, so strangers downloaded packages that shipped two major versions in a week. Packages were coupled without anyone deciding it, one merge touched 1,236 files, the home page overstated what the default install gives you, and a one-button page shipped 190 kB of JavaScript.
Worth knowing
The author concludes that agents will always call the work ready, and a phone is fine for spotting a broken screen but useless for checking a 300-commit merge. By the author's account, scope, architecture and release decisions were handed over far too early. The post now tells readers not to install the packages and to start a project with a skeleton and a reference to follow.