Agent teams

Inside Grok's bot-run sticker shop

A Grok Bot workshop connects shared design bots, competing website mock-ups and a proposed analytics routine, with human preparation and review throughout.

Give the bots separate jobs

Before Joseph can ask his bots to keep a sticker shop changing on its own, he has to work through their individual assignments and rehearse the complete workflow. In Grok's workshop, the SpaceX AI field engineer describes the iterations behind a proposed routine for finding product-market fit. "Human in the loop is still quite important", he says; the preparation includes working out the rough edges between bots as well as within each one's task.

The demonstration begins with a smaller delegation: Joseph asks Roy, his primary bot, to tell an X bot to announce that the session is live, then opens the X bot's computer to show it working in a signed-in account. Each bot has its own memory and computer, he explains, while the primary bot holds context about what the others are doing. Scott Lerner, who introduces himself as part of the AI deployment team, hosts the session as Joseph walks through the interface.

Sharing a bot lets Joseph reuse work that another person has configured. He brings Dex, a designer created and shared by Martin, into the sticker-shop project; Martin had already used it to turn Figma assets into creative designs with the Grok API for printed stickers, so Joseph reuses the shared designer for the shop.

Compare the same design assignment

The website brings Dex together with Lou, a full-stack developer bot, and Sam, a UI and UX designer bot. To demonstrate how he developed it, Joseph starts a fresh designer bot, supplies the Duolingo store as a visual reference and sticker assets from Martin, and asks it to launch three models through Cursor cloud agents.

Joseph gives multiple models the same design task so he can choose the result he likes best, while Lerner clarifies that the model choices apply to cloud agents within Cursor. When Grok Bot reports back, Joseph opens one of the agent links and shows its brief.md file, which he compares to plan mode.

The competing models work on the same design assignment, while the specialized bots have separate responsibilities. Asked why he does not give everything to one bot, Joseph points to their individual computers, context and memory. "Context is king in this era", he says, comparing the division of work to roles on a startup team. Lerner says that refining a bot around its role has worked best in his experience.

Set the scope before implementation

Development starts with a voice conversation about scope. Joseph asks the full-stack bot for a lightweight sticker-shop website on Vercel, then rejects its proposed Stripe checkout because he wants a mock-up without a database. "I don't even want to hook up to Supabase", he says. The bot responds with a front-end arrangement using images and a hardcoded product list, and Joseph asks it to create a repository, launch a cloud agent and return screenshots and a recording; the bot says it is spinning up the scaffolding agent.

Joseph can also open Lou's computer to inspect the terminal and request a local development server, either trying the application himself or asking the bot to exercise it and send back a recording. When the first design result arrives from Grok 4.7, he inspects its assets, mobile view and screen recording, noting that the bot made the recording.

Use analytics to choose replacements

Once the website exists, visitor activity supplies information for changing its selection. Joseph asks Ed, the metrics bot, to update a popularity canvas using Vercel analytics; he has encouraged the audience to visit so that the dashboard contains activity beyond his own clicks. He shows an earlier dashboard and later returns to the refreshed result, using page views and unique visitors to identify stickers for replacement.

In the proposed routine, Roy coordinates the handoffs: Ed reads the metrics, Dex designs replacements for the bottom three stickers, and Lou swaps them into the code and deploys the updated website. Joseph describes running the process every 24 hours, with random selection among tied candidates, then invokes the PMF skill and shows Roy messaging the other bots as work begins. Lerner suggests a checkpoint such as a daily brief identifying the least-performing stickers, and Joseph agrees with keeping a person involved.

That combination of continued work and human checkpoints also appears in a separate product announcement. claudeai on X describes a proactive agent in Claude Duet that keeps working when the user steps away and checks before acting on their behalf. The post and Grok's demonstration concern different products, but both describe agent work that continues alongside points of human involvement.

Prepare the workflow before expanding it

Email extends the discussion of what a bot can do with its own account. Lerner describes newly released functionality that lets a Grok Bot claim an email address for service sign-ups or meeting-related tasks, while poteto on X supplies a concrete bot email address when asking someone to have another bot send a message.

The closing questions return to how Joseph prepares an assignment before sending it to competing models. He first develops the workflow with a model that he knows, then broadens the comparison once he has a prompt that produces what he wants; an imprecise prompt, he says, can be "interpreted in different ways" and waste tokens. Across the shared designer, the competing mock-ups and the analytics routine, the shop connects separate bot responsibilities through assignments that Joseph configures and reviews.