From September 15 to 17, 2026, three xAI employees attempted to build a company from scratch, live, with about thirty Grok Bot agents doing everything. They did deliver a playable game on Thursday. But the list of what broke in the meantime, from matchmaking bugs to the SQL query that cut off registrations, says more than the result.
The event was called “Grok Bot Galaxy,” held at the Howard in San Francisco, with a free livestream each day from 8:30 a.m. to 6 p.m. Pacific Time. On stage were Matt Palmer, Lauren Tan, and Roshan Sadanani, three members of the team developing Grok Bot, xAI’s agent product. If you think of the Grok chatbot you encounter on X, this is not it: a “Bot” has a name, a role, and a cloud computer. All Bots on the same account share this same machine, which matters for what follows.
xAI drops the pop-up restaurant for a Grok Bot card game
The starting point was deliberately empty (a GitHub organization with no repositories, a fresh Slack and Notion, no product idea) and the company was given a name that served as a program, Ship by Thursday, deliver before Thursday. The official @bot account opened the first day with a laconic post, “This is Day 1.” Useful clarification because the announcements spoke of “72 hours”: this meant three streaming days of about nine hours each, not a continuous timer.
The first day thus started with a “pop-up OS,” in other words software to help a chef set up a temporary restaurant in San Francisco, with the team organizing the first pop-up themselves to learn the trade. The idea lasted an afternoon, because food permits and alcohol licenses diverted the concept toward a generic pop-up, then an art exhibition, then tech merchandising.
On the second day, the pop-up died and was replaced by an auto-battle game in the browser, where Grok Bot marketplace bots become cards battling in teams of three. On Thursday, the game was released under the name Thursday Arena, with a training mode without an account and ranked matches via an X login and an Elo-type ranking.
How many people tried it? According to the end-of-stream screens noted by Japanese observer Kuroto_Works, 1,902 players had logged in with their X account, for 4,884 training sessions and 6,546 public matches. These figures were not audited by xAI, however.
Grok Bot merges over 400 PRs in 3 days, then crashes the game live
The speed, however, was real. The first day dashboard showed 26 bots online, and the counter later exceeded thirty. At the end of the stream, screens showed more than 400 pull requests processed during the build, and 157 commits by Roshan Sadanani alone.
But this pace ultimately proved costly. On the third day, an autonomous fix pushed a bad SQL query on the database which this time contained real users. The game crashed and registrations remained blocked until the rollback.
According to the incident log compiled by the author of the grokbot-field-notes repository (an outside observer of xAI, co-founder of the nopaywall.io catalog, whose post also promotes a paid subscription), Lauren Tan was explaining that agents need to be reined in at the precise moment the incident occurred. The “autopilot” branch had merged its hundreds of PRs on a disposable project, and no one tightened safeguards when it faced a production database.
The same repository lists forty things that broke on air, some dating back to the first day. The chief of staff bot opened a pull request while the team was deliberately pushing to the main branch without PRs, because the rule existed only in humans’ minds. Matt Palmer had to dictate it aloud, “new rule, no pull requests.” The GitHub connector also requested new authorization mid-stream, and cloud agents waited 300 seconds for tests that ended in one minute.
Thursday Arena: bugs without tests to reveal them
Once real players arrived, bugs piled up. Matchmaking that matched players only against AIs while Elo scores kept changing, a ranking whose win counter didn’t match match history, a complete freeze when spending all gold, an ad banner covering shop buttons on mobile, and spam in the feedback form within hours. The ad auction meant to fund the game launched broken and was still broken at closing.
These problems were spotted by humans playing, not by the machine. For good reason, most tests had been deliberately removed during the prototype phase to speed things up.
The AdvisioTech report draws the same conclusion, since problems were only seen once humans faced the product, and it infers that the bottleneck is no longer in code production but in evaluating what was produced. On the revenue side, the counter shows zero, the only “theoretical first revenue” of Ship by Thursday having been a $1 auction test on a system that didn’t work.
Grok Bot routines: a task every 15 minutes runs 100 times a day
Parallel sessions of the event gave a different figure than the demos, and it concerns your bill. Several participants had programmed routines every 15 minutes because the task seemed urgent. Yet Lauren Tan had already written at the end of August that a 15-minute routine runs nearly 100 times a day and each execution consumes tokens. Her advice is simple: once an hour or a few times a day is enough.
Product limits go in the same direction, since an account caps at 50 bots and group conversations combined, and each bot at 50 routines. On the third day, Blake from the xAI team summed it up in a phrase reported by the repository: “You don’t need 45 bots.”
Three people and about thirty agents thus delivered a game that thousands tried. It’s not a company in the sense of revenue, nor is it an autonomous system, since the pivot to the game, the launch, and the rollback after the crash remained human decisions. The question remains whether the next demo of this kind will be played with real tests, or if the public will continue to write them…
