이미지: AI 생성 — METAL LAB
Summary
- Two days after Grok Bot opened in early beta, real-use reports began accumulating on X.
- A game developer created 74 pieces of card art in two hours and wired them into the game, while another user completed a reservation entirely by voice while walking.
- On the flip side, one user complained that a file stayed trapped inside the bot's own computer and never reached them, and Musk said he would fix the basic issues first.
Two Days' Worth of Reports
Two days have passed since Grok Bot opened in early beta. The official pitch is brief: the bot is an AI teammate that does real work on the user's behalf, logging directly into the tools the user already uses and returning with finished results.
That sounds impressive, but the pitch alone doesn't tell you much. Fortunately, actual usage reports are already piling up on X. Setting aside promotional posts, here's what people who actually used it have said.
First, the structure. Each bot gets its own cloud computer. Once a user lets it watch them perform a task, it saves that process as a routine and repeats it automatically at set times. Multiple bots can run simultaneously, exchanging messages and handing off work to each other inside group chats.
The origin story wasn't grand. Cursor engineer Lauren brought up the internal Slack bot "Benny." "I'm pretty lazy, and Cursor was getting flooded with bug reports. I wanted to fix them all." The idea started as a way to have an agent fix bugs automatically overnight, and questions poured in asking how it was built and whether others could build one too — which eventually led to Grok Bot.
A Game Developer's Two Hours, 74 Pieces of Art
The most detailed report came from game developer Danny Liemansetya, who had about a week of early access to test it.
His first task was replacing every placeholder image in his game with real art. Before heading to his art-generation tool's webpage, the bot first read through the codebase. After identifying what each asset was, it wrote a separate prompt for each one, generated the images, cropped them, cleaned them into transparent PNGs with backgrounds removed, and reconnected them into the game.
74 pieces of card art in two hours — work that used to be done one piece at a time by hand. He wrote that he has since handed off all art-related tasks to the bot.
Other use cases follow a similar pattern. He set it up so that pushing a build to GitHub automatically uploads it to itch.io. Feeding it a PRD document and hooking it up to the Figma MCP got him UX flows and wireframes. It dug through his inbox to find forgotten paid subscriptions and unsubscribed him from newsletters on his behalf — though he noted it missed a few.
Some tasks were slow at first. Game playtesting was sluggish initially, but speeding up came from having the bot build a dedicated skill that understood the game's rules.
"I Expected It to Break, But It Just Kept Running"
Matt Shumer, who tested it two weeks earlier, offered a different angle. He built a researcher bot and a writer bot, then assigned both to a "Chief of Staff" bot to manage on a project.
He assumed it would break down, and went to check — but it was just running smoothly. His summary: "an agent for everything, not just code." He pointed to the interface resembling a messenger chat window, each bot having its own designated job, and the system learning the user's habits over time as its strengths.
He did note one clear complaint: the model router. Users can't choose which model handles a task — the backend assigns it automatically.
When the assignment goes well, it's arguably the best design for general users; when it doesn't, it's frustrating for power users. He added the caveat that he'd heard it has improved a lot since he tested it.
A Reservation Completed by Voice in a Parking Lot
The most widely shared report came from Yunta Chai, which passed 3.5 million views and spread further after Elon Musk quoted it.
While walking from a parking lot to his car, he spoke to the bot in a mix of Chinese and English. The bot scanned his calendar, identified something that needed to be booked in advance, determined the best time to book it, went to the website, and completed the reservation. His comment was short: "Impressed."
There are also entirely different use cases. AI YouTuber Wes Roth is having the bot run a Pokémon game livestream. He set up a Chief of Staff bot to oversee coach, host, player, avatar, and safety bots, and picked Bulbasaur as the starter Pokémon. However, since the first dry run hasn't happened yet, it's unconfirmed whether the stream actually ran.
What Piled Up on the Other Side
Not everything was praise. One Chinese-speaking user wore himself out trying to receive a single file. Three agents each sent back a file path pointing to their own respective computers, and none of them opened on his machine. Repeating the request multiple times didn't help.
This is the design of each bot having its own separate computer turning directly into friction. It's like a coworker telling you, "That file's in the second drawer of my desk" — if the drawers are all locked separately, knowing the exact location doesn't let you open them yourself.
Musk acknowledged the issue. He wrote that after fixing the basic problems in the early beta and releasing Grok 4.6, he would expand the beta. Grok Bot is currently in early beta and is included with Grok Heavy and Cursor Ultra subscriptions.
Editor's Take
Looking at two days' worth of reports side by side, one common thread emerges among the successful cases: the output existed as a file. 74 pieces of art, an uploaded build, an organized list of subscriptions, a confirmed reservation — all things a person can verify as done or not done in one second.
By contrast, the friction case involved output that stayed locked inside the bot's own computer. Giving each agent its own machine has the upside of not occupying the user's own computer, but the downside of the output never actually reaching the user's hands. Switch between several agent tools and you keep hitting the same wall — not a capability problem, but a handoff problem.
So for teams testing this tool right now, a better sequence might be: in the first week, only assign tasks whose output lands as a file or a link — generating reports, converting assets, organizing lists, automating uploads. For tasks that involve logging into an account and sending something out — emails, payments, confirmed bookings — insert a human checkpoint.
Shumer's complaint about the model router fits the same pattern. Not being able to choose the model means there's no lever to pull when results go wrong, no way to trace why. For general users that's one less burden, but for teams trying to put this into real workflows, it becomes a point where root-cause analysis gets blocked.
Finally, there's something worth taking from the Benny story. This product didn't start from a grand vision — it started from a narrow annoyance: too many bug reports, and a wish to handle them overnight while sleeping. The same order of operations applies to teams here adopting agents. The first step isn't deciding what to have the bot do — it's finding the one task that repeats every week that nobody wants to do.
