Launching Thursday Arena and
a factory that fixes what players reportA technical account of Galaxy Day 3

Introductions and a free trial for new users

Matt opened with, "Day 3, and today is the day we ship the company." Matt covers developer experience, Lauren, known as Potato, covers Grok Bot, and Roshan covers product; all 3 are SpaceX AI employees. Overnight progress was reportedly substantial, because Lauren had set up the software factory the night before.

Before the main topic, they announced that any new user who downloads Grok Bot, creates an account, and sets up 1 recurring task within 10 minutes gets 1 month free (the top-tier plan, worth about $200). The screen showed a QR code and a countdown timer. Examples of recurring tasks included sorting email, watching GitHub issues, and checking Slack, and Lauren said examples would go up on X.

The factory built overnight

For viewers who just joined, Roshan explained that they're building a game company and launching the game today, and that the previous afternoon, they'd assembled a team of bots into a "software factory." The factory on Roshan's own screen is reportedly a copy based on a bot template Lauren shared.

According to Lauren, the mechanics were rethought with Roshan the night before. The first version was a 1-on-1 rock-paper-scissors-like game with scoring and a leaderboard, but it wasn't fun to play, so the plan was rebuilt drawing on mechanics from games they liked. The tools used were Grok Bot and cloud agents, plus Lauren's own skill set pstack, a plugin for both Cursor and Grok Bot.

Potato Mode's fully automatic mode

pstack has a skill called Potato Mode that teaches a bot rigor and how to direct other bots and agents. Typing "/potato-mode fully automate this plan" splits the plan into many small stages, and spins up an agent that implements small PRs and a verification agent that actually runs the app to "fuzz" it (clicking around to find bugs). Because verification is strict, in fully automatic mode PRs merge themselves.

Lauren said more than 100 PRs went in overnight. The number shifted mid-sentence, from 145 to 150 to 170, as Lauren read off the closed-PR count on screen. The work covered a UI rebuild, backend updates, security, and load readiness, and Lauren added, "hopefully it holds up."

The playtester bot, "Play"

Having just merged PR #124 (a UI rewrite), Roshan showed the bot-to-bot interaction on screen. A bot named "Play" handles playtesting, standing in for the QA team in place of the intern from the day before. It watches each PR, either on a recurring task or in response to events, and when it finds a PR with a green CI, it operates the UI to play through the whole game and sends back what should change before the merge.

Lauren said the word "factory" isn't a favorite, but it's a close enough metaphor for the automation, and that repetitive work like an email inbox, Slack, or GitHub PRs can be automated with Grok Bot's recurring tasks.

A factory is "a system for running agent iterations"

Matt restated the idea of the factory: build a good feedback loop where an agent can verify its own work, removing the manual labor of clicking through the UI to check things or reading logs; recognize that this applies not just to development but to email and business operations too; and break down a person's work to hand off the part an agent should do. Sending an email might not be an agent's job, but clearing spam, unsubscribing, and drafting can be.

The state of the game before launch

Roshan pulled up the game's screen. The working title is Cupcake; the real name will be revealed at launch. It has sign-in with X, an overall leaderboard (at this point just Lauren and Roshan), tiers like Platinum, and a guest mode added the night before that lets players try it without signing in. The front page has a sponsorship sign-up form, which they said would be turned on later.

To play, you pick a lineup from a pool of bots turned into cards, then get matched against other players over several rounds. At the start of each round you receive gold (a proposal came up to rename it tokens), which you spend at a shop to buy bots and upgrades and sell off bots you don't need. You start with 10 gold, and winning all 3 rounds wins the match. The bots are real bots from the Grok Bot marketplace, and they generated unique avatars for more than 70 of them. CSS was still broken and clipped in places, and Roshan said, "we know what needs fixing."

Splitting up the work until 9 a.m.

They laid out the plan to cut to the main-stage talk, "Grok Bot for Marketing Operations," at 9 a.m. Pacific, and afterward bring the speaker up to the studio to discuss the launch. The venue is a building half a block from Dreamforce in San Francisco, with the main stage on the lower floor and the streaming studio upstairs.

The work running in the background was Roshan on visual polish, Lauren on backend load testing that bots had hardened the night before, and Matt on a system for collecting requests and feedback, storing them safely, showing them to the team, and feeding them into implementation. The night before, a bot handling audio had made a track with Suno and another bot had prototyped 3D animation, but those stayed low priority.

The stream's on-screen schedule, and the login bug

This day's stream kept a schedule and a bot list displayed on the right of the screen at all times. Matt's chief-of-staff bot, Steve, reportedly runs the proceedings. Roshan revealed that Roshan's own chief-of-staff bot used to share the name Steve, and had been renamed the day before, drawing laughs over the candidate names.

Right before cutting to the talk, logging in with X errored out in Matt's environment. Matt guessed the automation was Lauren's doing, said the server side would be checked first before handing off to Grok Bot, and closed with, "we'll fix it within the hour."

The premise of Grok Bot, and "tools, not rules"

The talk was already underway by the time the stream cut over, covering how an always-on agent's advantage is that the work doesn't stop when you close your laptop. Matthew's introduction named RevOps and MarOps at SpaceX AI as the coverage area, and confirmed that many people in the room had the same kind of role.

If limited to just 1 use case out of many, it's that "you can build tools instead of rules." Instead of handing out a checklist of the correct steps for setting up a campaign and asking people to follow it, you build an internal app that's designed to follow the rules from the start. Sales and marketing, in this framing, become the players of a game the ops team designed.

4 bots named after music gear

The bots' names come from music gear. The chief of staff is the synthesizer OP-1, the inbox bot is Fisher, the product manager is Juno, and the engineer is Ondes, after the Ondes Martenot, which Matthew called a nod to Radiohead fans. Ondes knows the internal CRM, the data warehouse, the existing codebase, and how to build in-house tools.

A task list that clears itself

On screen, Fisher had finished its scheduled intake run at 9:04 a.m. and, during the talk, sent OP-1 a roundup of what was new in email, calendar, and work chat. Matthew explained that even with notifications off, a bot watching in the background surfaces only what matters, cutting down on context switching.

OP-1 carves the actual tasks and subtasks out of what it receives. It reads an entire Slack thread and restates "what is this person actually asking for," so it can handle messages you couldn't judge from a single read.

Handing a territory dispute to a specialist bot

The "clears itself" part comes from OP-1 relaying tasks it finds to another bot. In the example, 2 AEs were arguing over whose account territory it was, and OP-1 passed it, with the background attached, to the territory-planning bot. That bot has the rules of engagement and the CRM context, and drafted the first pass of what was needed.

Matthew said the dispute needed no thinking through in advance, just checking and finishing the content afterward. The division of labor is inbox monitoring, task extraction, relaying to the right specialist, and final oversight.

A matching app for leads

The second demo was a "matching app for leads" used inside SpaceX AI. To fix the classic handoff problem, where sales lets leads marketing gathered sit idle and no feedback flows back to targeting, each AE logs in and sees only their own leads; swiping left picks a reason and rejects the lead, writing it back to the CRM correctly, and swiping right enters it into a follow-up sequence.

Instead of an ops person pushing for SLA compliance, Matthew explained the goal as building a game for sales to play, using the rejection-reason data to improve marketing's targeting.

Juno asks before it builds

The build starts by asking Juno for "an internal app where sales reps review leads," along with a rough mechanism, and telling it that writing back to the CRM is the top requirement. An older AI tool would have tried to please the person asking and started writing code right away, but Juno asks follow-up questions instead.

What Matthew answered on the spot: show only new leads with no extra filters, don't change CRM fields on accept (status changes once the rep makes contact, and existing automation records that), attach a rejection reason and an optional note, queue offline actions for review, and keep it local only. For Matthew, answering the questions was what deepened the thinking behind the app.

An automatic handoff from spec to Ondes

Juno wrote a product spec and, without being asked, passed it to the engineer Ondes to start implementation. Matthew called this the key part: before, even getting a good output meant carrying it into a separate system by hand.

The finished product was shown cooking-show style. It's all fictional data: dragging a lead card left rejects it, dragging it right enters a sequence, and the lead's source is also visible. It reportedly took 2 weeks from idea to implementation, with the actual build taking about 10 hours, and the rest spent rolling it out to sales reps and gathering feedback. The lead-review rate "shot up" over those 2 weeks, Matthew said, though no figures were given.

3 takeaways

"If you could hire 1 team of direct reports, who would you hire" is the starting point for assembling bots, Matthew said, and Grok Bot opens up to everyone the "CEO's life" of being able to hire an unlimited number of specialists with the right context and authority.

Second, trust the bots with authority while also setting guardrails. Knowing what authority a bot has lets you hand off bigger work, and because sending a message, writing to the CRM, and submitting a PR all come back to the person first, nothing runs away. Last, Matthew told the room to think of the ops role not as the person who decides rules and procedures, but as a product manager whose product is revenue growth. Referencing Dreamforce, running the same week, Matthew summed up the daily job as finding and removing growth bottlenecks.

Q&A: WhatsApp, ad integrations, 1Password

Asked about wanting to use WhatsApp with Grok Bot, Matthew's answer was that no native connection was known, but that Beeper, which consolidates several messaging apps, had been put directly on Matthew's own bot's computer. Since any app can be installed on a bot's computer, a WhatsApp app should work the same way.

Asked to connect Meta's and Google's ad platforms directly and remove the problem of logins expiring, Matthew said the roadmap couldn't be discussed, but pointed to the 1Password integration announced the day before. The explanation was that a vault can be shared with a bot, and a routine can be set up to periodically check whether a login is still valid and re-authenticate if needed.

To automate a complex piece of work, Matthew recommended breaking it into basic chunks, logging into a platform and viewing analytics, pulling a report, reading it alongside other data, and building a specialist bot for each chunk that then work together.

Q&A: disappearing tools, multi-person bots, human work

Asked which marketing tool disappears first and what the relationship between bots and people looks like, Matthew answered that a bot is a mirror, and having questions asked back leads to deeper thinking. Matthew avoided a firm claim about tools, saying that the same thing happening in coding, where the barrier to entry drops and an engineer's value shifts to higher-level thinking and intuition for systems, will happen in other roles too.

On how 2 people's Grok Bots collaborate on the same project, Matthew answered, "I can't get ahead of the company, but there's an interesting announcement coming soon," and declined to answer about deployment on Tesla. Juno's template could reportedly be shared after the stream. The sales-side reaction has been positive, "building your pipeline while swiping on the subway," and for the last question, "what work should people keep," Matthew answered: building relationships, earning trust, and solving problems.

The challenge to travel to Starbase

To close, Matthew announced a challenge that had started on X. Introduce a Grok Bot you can't live without, and you and 1 guest can go watch a Starship launch at Starbase in Texas. To enter, follow @Grok and @Bot, and quote the challenge post with a description of your bot and a link to a shared template. A QR code appeared on the slide.

No screen while handling secrets

Roshan opened with, "this is where deployment happens." Because client secret keys and similar things are being handled, the screen stays hidden so production secrets don't appear on stream. Roshan described the game as a card-based battle game in the middle of a visual polish pass, and added, "please keep your expectations low."

For the day's plan, Roshan said they'd move between the downstairs talks and the upstairs studio, and in the afternoon bring in a SpaceX AI colleague to discuss monetization and help with marketing and revenue plans. First they'd release it as a playtest, then keep shipping and fixing after that.

What the factory really is, and the tools in use

A Grok Bot team manages the software factory. When it detects a bug in production, a bot asks a cloud agent for a fix and a PR, and the playtester bot plays through the change in the game before approving it, on top of CI approval. All of this runs on the combination of Grok Bot and GitHub.

Auth runs on Clerk, and a developer account was set up for logging in with X. A logged-in user gets an Elo score and appears on the leaderboard. There's also a practice mode. Matt was sorting out the Clerk config, where the domain points, and the DNS records, and joked, "it's always DNS," "and then CORS." Hosting is on Vercel, and the playtester bot has been using the per-PR preview URL. Moving to production means tightening database connections, input/output controls, and security. The domain name is decided but stays hidden until launch.

PlanetScale, and triage on Slack

The database is PlanetScale, and they'd been load-testing the several databases holding game state since the night before, and continued hardening them for production that morning.

On the operations side, there's a Slack workspace, and reports had started coming into the bug-report channel. Inside SpaceX AI, they put a triage bot on each product's feedback-and-issues channel, and it checks whether a report reproduces, looks at internal logs, and returns whether a fix is needed and who should be notified. That shrinks the time from report to fix and cuts down on people digging through logs. They plan to build the same system for this game too, but have it return a PR instead of findings or a dashboard.

Production is "plumbing work"

The UI rewrite and the auth overhaul were happening at the same time. Switching from development to production throws up a lot of "plumbing" work, like reconfiguring auth callbacks. Roshan said that most of building a company, and building software, is connecting things, and that agents are well suited to diagnosis that used to be manual work.

As a way of thinking about debugging, Roshan used the example that when a lamp won't turn on, you check the bulb and the plug first, and if both are fine, the cause is somewhere in between. The principle of enumerating and checking every point that could break between point A and point B holds for software and for business alike.

The first users are logged in production

Matt logged into the Clerk dashboard, without showing it on screen, and reported that it showed 1 production user, then 3. It was right after the switch from development to production, the moment auth and the game came together.

The Notion launch checklist

Roshan switched screens and opened Notion. The work board the agents had written to the day before was messy, but it held a to-do list, the canonical database schema, the source of truth for the game's mechanics, and other documents the bots use to share company context with each other. Roshan cited a point Lauren made on day 1, that a codebase is the best context for a coding agent, but operational matters are better written down.

Roshan created a new launch checklist and checked off the first item, "auth works." Whether to land the UI rewrite first was settled as "not needed," and the plan became to launch with the first playtest under the stated premise that the look isn't finished and it's likely to break.

Hide the URL, play through it, then share

Lauren proposed, "let's hide the browser URL, play it ourselves, and if it's fine, post the URL in chat and call it the first playtest." The site was already live, but the name change needed to be finished first. On Roshan's screen, the game came up, with Roshan saying, "moment of truth."

How to play, in practice mode

Roshan entered practice mode, responding to Lauren's request, "there have been so many changes it's hard to follow, walk us through how to play." Cards are built from Grok Bot templates and carry marketplace names such as Office Ops Desk, Writing Bot, and Dial Bot. Each card has 2 stats, attack and HP, plus an ability built specifically for that card, and the ability is what changes the game.

The game runs 3 rounds, and each round you set your bot lineup and battle an opponent who has also set theirs. You shop between rounds using your remaining gold. A basic card costs 3 gold, and you start with 10.

Combining abilities

Office Ops Desk's "Hold the line" heals 1 HP when hit from the front. Dial Bot's "Poke" deals 1 damage to the enemy's front card at the start of battle. Writing Bot boosts the attack of the ally in front of it. Roshan put Writing Bot in second position to strengthen the front Office Ops Desk, building a combo that opens with 4 attack.

Along the way, Roshan found a bug where a card's name turns into "unknown bot" right after purchase, and flagged it for a fix. When someone pointed out the on-screen text was small, the response was to pinch-zoom in.

Round 1 and Round 2

Between rounds you can buy a +1/+1 upgrade for 3 gold, but there wasn't enough gold, so Roshan fought as-is and won Round 1. Battles display abilities as a stream of events, but someone said it moved too fast to follow, so a step-through option and an autoplay button were both considered, and the decision was to keep the MVP fast and just add autoplay.

Before Round 2, the shop offered the uncommon Nightly Audit Engineer, which gains +2 attack when the card in front of it falls, and SEO and AEO, which deals 1 damage to the enemy in the matching slot. Roshan sold Dial Bot for 1 gold (flagging that the sell price should display), bought Nightly Audit Engineer, and stacked an Apple upgrade on the front card to make it a 4/6. The practice opponent is an AI that's "not that smart," but after launch, matches run against real players with tracked scores.

The bot list and rarity

The front page has a bot list, which Lauren called "a bug, this shouldn't be showing," and explained that every featured template on x.ai/bot/marketplace had been converted into a game bot with a generated avatar for each. Rarities are common, uncommon, rare, and epic, and the examples raised were Lingxi's bot and Dr. Eggbot, whose rarity got "cheated" up to epic.

An idea came up to skip making legendary bots of the team's own, and instead add a viewer-shared bot template to the game as legendary if they liked it. A template can be exported from Grok Bot's share button.

A 5-second discussion, and the first post

The discussion over "should we ship it now" took 5 seconds. Matt held the X account, which was still empty, but the URL was ready. They decided to ship it, clearly stated as a playtest expected to break, with Matt handling the first post.

For viewers who'd just joined, Roshan explained this was day 3 of a 72-hour stream, that they'd been building the idea settled on between the end of day 1 and the start of day 2, and stressed that the domain and X account they were launching are the "one and only real source," and that no other account or affiliation is theirs.

The name and account for Thursday Arena

The moment the post went up, the screen cut to Matt's, and @thursdayarena went public with its default profile picture still in place. Roshan repeated, "the @thursdayarena X account and thursdayarena.com are the only official sources; anything from anywhere else isn't us." Lauren reposted it and followed the account.

Matt pulled a cupcake emoji from a site Matt built for "extracting Apple emoji as PNGs" and made it the first profile picture. In parallel they were working on a favicon, styling, and an OG image, and Matt acknowledged the metadata that controls how search results and links appear wasn't finished yet.

The leaderboard fills up

About 2 minutes after the post, 96 people were on the leaderboard. There was no pagination, and sign-in was working. Roshan was in first place but started slipping right away, while Lauren, saying "I was stuck in second for 24 hours," kept dropping too.

As a standard step Roshan takes at every product launch, a data-scientist bot was asked for a "launch pulse every 15 minutes," delivering a dashboard with new signups, practice rounds played, and estimated conversion from practice accounts to real accounts.

The rankings shuffle, and followers climb

On Lauren's screen, a viewer in second place was at 1046 points to Roshan's 1047, a close race that a viewer soon won outright, taking first. X followers reached 225, and ideas came up for refreshing the leaderboard every minute and giving prizes to the top players.

Lauren said the factory had been auto-merging PRs overnight, and that Lauren's only input, "without looking at a single line of code," was asking Potato Mode and pstack to "make sure it can handle a swarm coming in from X." In the commit list, most of the commits in the previous hour were Roshan's.

The backend setup, and type validation with Zod

Users were approaching 500. The backend is written in Go and runs on Vercel serverless functions, with PlanetScale as the database. The client and backend talk over REST.

Lauren raised type safety as a challenge when 3 people are touching the code at once. Remove a field from the backend's Bot type, and it breaks if the front end is still reading that field. With more time, a different design might have been chosen, but for now the backend's response gets parsed and validated with Zod. A QR code for thursdayarena.com appeared on screen.

A dashboard that arrives inside the conversation

Roshan says "a business without a dashboard isn't a business," and while a founding engineer bot inside Grok Bot watches over PRs, Roshan has a data bot pull the latest metrics. Instead of logging into a dashboard site and clicking through screens, Roshan gets the exact cut of data needed right there.

Practice sessions pass 600, with 38 conversions from practice to signing in with X, 400 sign-ins total, and traffic shaped like a hockey stick. There are also signed-in match counts by stage (store, results, battle), which show how well the "sign in to continue" buttons placed around the product are working.

The metrics that matter before monetization

When Lauren asks "what does a product manager look at to set priorities," Roshan says it depends on the product's stage. Right now, in the zero-to-one stage with no monetization, the team watches funnel metrics: whether people are playing, whether they're enjoying it, and whether they finish the game and come back to play again. As guests on Day 1 and Day 2 kept repeating, revenue will eventually become the main metric.

As a game, they watch variation, replayability, balance so that no one keeps winning or losing streaks, and competitive dynamics. An idea also came up to give a richer experience or perks only to people who sign in.

The losing streak and the "only matched against AI" problem

The win/loss tally the bot produced shows a 41.8% loss-heavy split, with opponent types split into "ghost player, AI, unknown" (names to be improved). Lauren explains that battles are run against a snapshot of another player's lineup, and during testing there were no opponents, so it defaulted to AI, but Lauren notices that even after launch they only ever get matched against AI and their Elo never changes. On the leaderboard, viewers were starting to reach Platinum.

Roshan says a competitive game's win rate should aim for around 50%, and asks their founding engineer bot by voice: "Something's wrong with Elo or matchmaking, since it's always matching against AI. Investigate with Potato Mode and Cursor's cloud agents, don't open a PR yet, and report the cause." Lauren says the most important thing about Grok Bot isn't adding more bots but connecting every tool and plugin, and shows their own connections (pstack, Slack, Vercel, X, PlanetScale, Notion, Clerk, Excalidraw). Pointing at an MCP URL lets the bot set up the connection itself, and 1Password was added the day before.

In-game feedback and Slack

Matt shows on their own screen how text written through the in-game "send feedback" feature arrives in a Slack channel. Since it's accepting content written by users, it runs rate limiting, profanity filtering, and sanitization behind the scenes before writing to the database, and it's placed only on a path that has passed authentication with X. As a next step, Matt raises the idea of having Grok Bot summarize the feedback and sort it by feature so the product team can absorb it.

According to a later explanation, the client limits the length, the server strips profanity and sanitizes against injection attacks, and it further runs through an xAI model for a final check before rate limiting is applied. Some users got 429 (too many requests) responses, and the API was somewhat overloaded.

Asking Steve to build a "feedback factory"

Lauren hands their chief of staff, Steve, the Slack channel link and asks by voice to "build a factory pipeline that looks at user feedback, classifies it, tries to reproduce it, and creates tickets in the Notion database." Lauren puts off the stage of fixing it automatically and opening a PR, includes instructions to exclude content unrelated to the game, profanity, and suspicious links, to watch for prompt injection if AI is reviewing the feedback, and finishes with "say it back in your own words."

Around the same time, the first 5 or 6 Platinum players appear on the leaderboard, and an idea comes up (pending approval) to gift the top 10 players a Grok Bot Ultra plan. Roshan raises an issue with the store's design: since gold resets to the same amount every round, players end up repeating "sell 1, buy 1, upgrade" every time.

The classification result: 71% bugs

On Matt's screen, feedback lines up in the "Cupcake feedback" channel: a mix of requests for extra gold purchases, theme requests, and bug reports. Lauren asks their newly built bot to classify it and post a chart, and 3 items rise to the top: mobile layout breakage, only getting matched against AI with no ranked mode, and Elo. Bugs made up 71%, praise 16%.

Roshan says that as a game company, bugs, stability, and performance should come before new features. Roshan also reports smaller issues Matt found, like a bug where spending all your gold leaves you stuck unable to buy anything, and the missing logout button (there is a sign-out).

OG images and the favicon

Matt merges a PR that adds a favicon and meta tags, and shows the OG tags appearing in the developer console and how the link looks when pasted into X (the v1 OG image). Matt walks through checking it by pasting the URL into an OG preview site, and says the first step to being found, by people and eventually by agents, is "telling them exactly what the product is." X followers had reached 476.

Dividing the work between Crumble, Tater, and Hashbrown

Through Steve, Lauren has Dr. Eggbot build "a bot that classifies feedback, reproduces it together with a playtester role, and files tickets in Notion once confirmed." The new bot is named Crumble, and it reads feedback and finds P0 issues: a blank white screen, an empty board, and stale ranking and Elo displays. Matt invites Cursor into the Slack channel.

Lauren then writes that confirmed bugs should go to Tater to fix, Hashbrown should verify that the classification correctly understands the user's intent, and once that's done, they can move to fully automatic fixes and a "verification swarm." To avoid breaking the game for everyone in production, they always verify with pstack's /verify-cupcake skill and the swarm skill, and only fix bugs that could first be reproduced. The swarm skill spins up many Cursor cloud agents, each running the game on its own computer and clicking through it (internally called "fuzzing").

The outer loop and the inner loop

Lauren locates the reason for connecting all the external data, Vercel, PlanetScale, Slack feedback, right here. An outer loop that gathers and understands incoming data feeds the inner loop of engineering. When it works well, bugs get fixed automatically, the game needs less human oversight, and they get to spend their time on features and polish. Roshan adds, "spending time building the plumbing between agents pays off big later."

As a wrap-up just after 11 a.m. Pacific time, they repeat that the only official sources are thursdayarena.com and @thursdayarena, announce that growth lead Vincent will join at 11:15, and cut to the same Nokia customer footage as the day before.

What growth means

Roshan (@roshan_s on X) pivots away from Tuesday's physical pop-up idea, looks back on launching Thursday Arena that morning, and introduces Vincent. Vincent is on the growth team, whose job is reaching people interested in the product and helping them find how to use it. The shape of that job varies by company: in a product-led business like Grok Bot, they watch user behavior and push from inside the product, while in sales-led growth, email, outbound outreach, and LinkedIn are central.

Raising ideas while watching the funnel

Roshan shows the flow from logging in with X to the leaderboard. Vincent's first idea is to show a sharing pop-up the moment someone sees the leaderboard, so they can brag about their rank on X, like an old arcade. Roshan creates a bot called "Vincent growth" on the spot and asks it to build a growth playbook in Notion to hold the ideas. The leaderboard had nearly 1,000 people on it.

Explaining the game loop, Vincent praises the fact that cards are marketplace bots as "growth material," and an idea comes up to let people bring their own bots into the game. Players think about lineup order around abilities like: when this dies, give +1/+1 to the 2 units behind it; deal 1 damage to the enemy in the back at the start; or pass 50% of attack power to the ally in front. A card for Critiquito, made by design team member Manuel, also appears.

Bringing sharing and rarity into the growth loop

Vincent's second idea uses rarity as a growth mechanic: you share a bot or card you made, and if someone else adopts it, you get a rarer version of your own. Other ideas follow: skipping the store to save up several rounds' worth of gold, putting a share button on the victory screen, and letting players name an opponent and challenge them directly (requiring the other person's consent, similar to Words with Friends).

Roshan explains that they just hand ideas over by voice, and the bot organizes them in Notion from there, while factory agents try them with PRs and playtests. The home page that appears after the victory screen still had things to fix.

Automatic DMs to followers and mentions on X

Vincent praises having owned touchpoints, the site and the X account, and says that now, with many concurrent viewers, is exactly the time to narrow down the single "action you want people to take." The idea: automatically DM anyone who follows @thursdayarena, thanking them, pointing them to the game, and giving a perk like 10 gold. It's also a countermeasure for the fact that traffic right now is limited to the stream and X posts.

Next comes the idea of using the X API to collect posts about Thursday Arena, reactions, criticism, reviews. Since Grok Bot already has an X connector wired up, Roshan asks it on the spot to "search for feedback on X and summarize it with handles in a new section in Notion."

The funnel from practice to sign-in

The data bot's latest figures (11:25 a.m. Pacific time) show about 2,200 practice rounds and about 1,000 sign-ins with X. Vincent suggests keeping the design that allows practicing without signing in, but showing "sign in with X to continue" after 2 or 3 practice rounds, and, citing a roommate who works in the game industry, making practice easier to win so players feel like they "want to play."

Roshan admits that the win/loss tally right after launch showed a fairly hard game, that practice and post-login should be looked at separately, and that there's no difficulty tiering yet. Roshan reveals that a game design bot named Krit had criticized it as "too hard at launch." New sign-ups had settled down after the initial spike right after launch.

Outbound email and the recipient's perspective

On outbound email, Vincent says recipients look at the sender's name before the subject line, so it should come from someone recognizable from X or LinkedIn. As a joke, sending email under Lauren's (potato's) name comes up. Ideas also include experimenting with giving a bot its own email address, and building a contact list by pulling people like game company developer-community leads off LinkedIn.

Monetization patterns and storing player value

Acknowledging the "gather users, monetize later" stance and asking for thoughts anyway, Vincent, prefacing that they aren't a game expert, cites the cooldowns-and-ads pattern in free match-3 games like Candy Crush, and paying to shorten wait times as in Clash of Clans. Vincent says the key is finding what makes players want more, and monetization inevitably carries friction.

Roshan shows the feedback form on the site. Right after a deploy, they're asked to log in again, and the form now has a phone number field (Matt's addition). Vincent proposes giving gold for sending feedback or bug reports, and Roshan sums it up as "storing value that makes a player invested and want to come back."

Sorting the growth playbook

The Notion growth playbook has been formatted by a bot, with impact and effort scores attached to each idea. The list includes sharing a win, sharing a card, challenge requests, weekly clips or leaderboard images, avatar showcases, a welcome DM after a follow, a pop-up after 2 or 3 practice rounds, and a mailing list. Vincent pushes the follow DM as "speed is everything," picks the post-practice sign-in prompt as "high impact, low effort," and picks sharing the leaderboard as something worth being proud of "precisely because the game is hard."

The top 3 go to the engineering factory. Vincent raises a caution, "don't overdo it and chase users around," and finally recommends nailing the game's PMF first, then thinking of marketing (awareness) and growth (maximizing the experience for people who already came in) as separate.

How Vincent uses it personally: sending bots to courses, standups

Growth work is understaffed, and 1 person often plays data scientist, designer, and email lead at once. The night before, Vincent asked a new bot to "take the 101, 201, and 301 courses to become the best data scientist," giving it good traits as lasting skills. They did the same for their designer bot. They also teach basics from their banking days, like "use formulas instead of hardcoding anything but inputs."

Vincent's favorite use is having their main bot run a standup for all the bots every morning at 9 and put out a summary every evening, a hybrid of 1 main bot plus purpose-built bots. Roshan responds with a comparison to onboarding a new hire and teaching them how the work gets done.

2 perks and a break

2 perks are announced: existing users can type "reset" in chat to get their usage limit reset the same day (or the next), and new users who create an account at x.ai/bot and set up a recurring task within 15 minutes get a month free. Vincent describes growth as "winning people's hearts" and asking "how would I want to be treated if I were the user."

They repeat that the official sources are thursdayarena.com (no www) and @thursdayarena (Roshan misspoke once and said "on Slack"), show Vincent's X handle on screen, and go into a 2-minute break before the next guest.

For colleagues, "later" is now

Roshan welcomes Matthew, the morning speaker, into the studio. Matthew is on the marketing operations team, building tools and automation so that marketing and sales teams work efficiently, and enabling those teams to use Grok Bot. To the joke "gather users, monetize later," Matthew's colleagues replied that "later has already arrived."

Handing every inbox to 1 bot

Matthew's 2 main use cases were staying sane and cutting context-switching. A chief-of-staff bot is a pattern many people adopt, but the key was handing that bot every inbox, multiple email accounts, calendars, chat, Slack. It regularly receives everything coming in, figures out what's actually a task, and picks the specialist bot best suited to draft the first response.

An example is a territory dispute: a sales rep met someone at an event who turned out to belong to a colleague's assigned account. Judging that against the rules and procedures used to be a human call, but now it can automatically show alignment with the rules of engagement and what to do next.

What operations means for a 3-person company

Asked "does a 3-person company with 1 game on a website need sales or marketing operations," Matthew returns to a principle about business itself. A business exists to connect with people and deliver value, fun, or memories. For a consumer product, that means reaching the people who'll enjoy it most, and keeping the ship moving comes down to either paid access or advertising.

Showing Matthew the game and the marketplace

Roshan logs in and shows the game. There are 9 Platinum players, they're ranked 37th, over 1,000 users have logged in, and some display glitches show up too. The home page has a sponsorship sign-up form. Matthew looks for their own marketplace bot, "Inbot," but it's nowhere in the game (it may have been renamed).

Instead, they check the description of "Cooper" on x.ai/bot/marketplace. Published by Tommy Hansen, it covers AI, tech, venture, and business news, delivering a summary to Slack every morning at 8 and flagging competitor moves during weekday hours. Marketplace bots come with memory (role context), recurring tasks, and connections, and Matthew says, "unlike a tool that's so complex you can only copy it, I use it as a seed for ideas, and I can build the same quality myself."

Importing Cerebro

The template Matthew shared beforehand was Cerebro. Named after the scene in the X-Men movies where Professor X uses a device to find mutants around the world, it's a tool for thinking through "what you built, who you built it for, and how to find those people" at the level of individual humans rather than industries or revenue. It sets an ideal customer profile (ICP) and thinks through the data you need.

When Roshan imports it, Clay MCP and Amplemarket MCP handlers come attached as context, and they put off connecting them. Roshan pins Cerebro, labels it GTM, and asks it to get a summary of the game from their chief of staff and Dr. Eggbot.

The ideal customer profile for ad slots

Matthew says avoiding pay-to-win means leaning on advertising instead, but letting in anyone who'll pay makes the experience inconsistent. What they ask Cerebro to do is work out the ideal profile of who should fill the ad slots, adding off-the-cuff ideas like people who want to browse other games while playing, or people who want a snack. Matthew's own "Find My ICP" skill wasn't in the template and couldn't be found as a slash command, but Cerebro pulled up the procedure from memory.

The output that comes back lines up ICP, personas, segments, scoring, and signal recipes, with candidates including other games, indie studios, casual games, game-discovery platforms, snacks, energy drinks, and merchandise. Matthew explains each term. ICP is "whoever shows up on the device's screen," personas are the main categories of counterpart (brands, sponsors, platforms that send traffic), and segments are those bundled by priority.

2 kinds of growth: the PLG and SLG flywheel

Roshan looks back on the previous day's monetization discussion (avoiding pay-to-win, selling screen space as ads) and says they hadn't considered the angle of growing traffic through partnerships. Matthew splits growth into 2: consumer-facing PLG (getting more people to play) and sales-led SLG (outreach to a small number of partners and ad-slot buyers), where SLG's cold emails cite PLG's numbers. Partners bring traffic, and as traffic grows, the value you can offer the next partner grows too, a flywheel.

A signal recipe is information that sharpens the outline of a target: narrowing it down to, say, "a recently founded snack company that makes a slightly healthier version of junk food," makes the ad feel natural to that target and gives them a reason to reply.

Pulling numbers straight from the bot

The metrics bot shows over 3,000 matches in a few hours, a spike right after launch, and about 600 sign-ups, and it had changed the chart colors on its own. Roshan asks it to add hourly and cumulative user counts to the dashboard.

There's still no hypothesis about "who's actually playing," so Roshan comes up with the idea of people who want light mental play during a spare 5 minutes, a crowd like Words with Friends, and dictates that straight to Cerebro. Cerebro asks other bots for email material and returns a draft opening: "launched this morning, 1,000 players, 440 matches in the last hour." Roshan says this used to be work the partnerships team would have to request from the data team.

Updated segments and a line that reads "AI-flavored"

The segments update to "casuals" and "scroll substitutes" (a stand-in for short-form video), lining up personas and candidate partners: news apps, a newsletter like Morning Brew, a stock-checking app for retail investors, and physical-movement breaks. Looking at a one-line pitch, "scrolling the news, a snack, and 1 more match, all sitting in the same 5 minutes. Stadium ads for a brand that already owns that break," Roshan laughs that it "reads a little AI, there's even an em dash in there," and says they'll use it as raw material before a human edits it.

The Clay and Amplemarket flow

Since the connections weren't set up, this stayed at the level of explaining the flow. Clay MCP would search its database for companies and people matching the chosen persona, first surfacing 10 companies to approach about ads or partnerships. Next it would surface a contact at each company, and Amplemarket, the email tool, would send a personalized email referencing specific facts about that company. Grok Bot would watch the inbox for replies. Matthew adds that without a live connection, the bot might just start researching over web search instead.

Roshan says they'd like to set up Clay and Amplemarket accounts outside the stream and actually try running a partnership motion. After a noon wrap-up, the broadcast cuts to Blake Schuller's main-stage talk aimed at customer success.

Stages of AI maturity and a Grok Bot overview

Blake is an AI adoption manager, helping customers get results from SpaceX AI's product after a contract closes. The talk is structured as an introduction to Grok Bot, post-sales use cases, a demo, lessons learned, Q&A, and building live in the room.

Blake lays out a maturity progression, "from chat, to copilot, to a bot, to a team of bots," and describes Grok Bot as having a screen that looks like iMessage, backed by a team of agents with skills, MCP, APIs, and their own computers. It runs 24 hours a day, so you don't have to walk around with your laptop open.

The 6 use cases

Blake's day is filled with email, Slack, Teams, and meetings. There were 6 use cases. A morning status board (overnight fires, today's meetings, anything they said they'd do yesterday and didn't). Prep that arrives 15 to 20 minutes before a call (attendees and their titles, what was discussed last time, new features shipped since, proposals, blockers to clear). A post-call workbench (the moment a call ends, it grabs the Granola transcript, pulls together context, and drafts replies and documents).

Tracking commitments (following through on what they said they'd do), tracking requests (making sure something they asked someone else to do doesn't fall through the cracks), and pulling account status on demand. For that last one, asking "what's the status on Harbor" gets, after 1 space joke, a summary of risk, stakeholders, blockers, unfinished commitments, Slack activity, and the next move.

Talking only to Gus

Gus is Blake's chief of staff, their "best friend." Many people talk directly to specialist bots, but Blake gets distracted easily, so they talk to only 1, and Gus manages a team of 10. "On any given day, I'm managing 1 or 2 of them," Blake says.

Frankie handles follow-up. Wally knows Blake's own writing voice (they've fed it past emails and Slack messages: casual, lowercase, and emoji-heavy for internal Slack, short and formal for an email to an executive). Trudy is a sourced source of truth that Gus queries whenever it needs accurate information. Scout watches 30 to 40 internal Slack channels and email and delivers the key points every day. Blake also builds a specialist bot per account, recommending 1 bot per customer for a small-to-midsize customer count (1,500 accounts would need a different approach).

Gus sits in on an internal meeting

On the Grok Bot screen, the left side has bots, pins, and a marketplace (other people's bots, plus plugins like Gmail, Google Calendar, and Slack), and the right side shows each bot's own computer. A "teach me a task" button lets you demonstrate an action for the bot to learn. The company used in the demo is a fictional airline, FlyLo.

Asked Gus to "sit in on the internal training session for me and send me the key points when it's done," it joins Google Meet, mutes itself, and posts in the chat, "This is Gus, Blake's bot." Blake says they don't send it into every meeting, they're selective about which ones.

The post-call bundle: email, Slack, an ROI document

Telling it "the Harbor call just ended" (this usually kicks off automatically), Gus contacts Wally, Frankie, and the Harbor bot, and returns "the Harbor post-call bundle, drafts only." Gmail has a draft to Maya and Priya, written in Blake's own exclamation-heavy voice. Slack has a draft to Alex, the sales rep on this account: "meeting the new champion next week, ROI PDF in progress." The ROI one-pager requested during the call was also built by Frankie in the FlyLo brand. Blake says this used to be 45 minutes of work.

Before the demo, Gus had also sent the decisions from the meeting it attended, new branding and links, and who owned which request.

Building a form in a VM, and pulling status on demand

Asking Franny directly to "build a Northwind ROI form," it opens Google Forms in a VM and starts entering questions. How fast VM work goes depends on the task, but SpaceX AI puts weight on how well the model handles VM work. Bots can be labeled in their description (chief of staff, and so on).

Asked "what's the status on Harbor," Gus queries the Harbor bot, Frankie, and Scout, and after a joke, "how do you plan a space party? You planet," it summarizes risk, stakeholders, blockers, unfinished commitments, Slack activity, and the next move. Specialist bots never notify Blake directly, only Gus does, with a blue marker.

Letting bots argue over time at a staff meeting

Calling a staff meeting on the agenda "what should I do with the only free hour I have today," several bots show up. Blake's standing instruction is "always push back, fight over my time." The Harbor bot raises an SSO issue, the Northwind bot says "usage dipped a little but it's fine, push it back," the Brightline bot also says it doesn't need anything, Frankie argues "respect the deadline, send it now," and Scout agrees.

The conclusion was SSO first, then the already-drafted next-steps email to Maya, and everything else on hold. Franny's form was finished, returning a share link, an edit link, and the questions themselves: team and role, primary use case, and the definition of done.

Starting from a voice memo

To deal with the fear of a blank page, Blake recommends a 10-to-15-minute voice memo. Say your name, your job, what you like about it, what's frustrating you right now, and which processes are broken, then hand the transcript to your first bot and ask it to "build a system that fits me." What comes back are suggestions like blocking time on the calendar for customer calls, or taking Slack monitoring off their plate. If it gets stuck, just say "don't stop until it's solved."

Turning it into a system: recurring tasks and a self-improvement scan

On the right side of the screen, recurring tasks include an 8:30 daily summary, unfinished commitments at 9 and 1, and a Friday dashboard. Blake's favorite is the self-improvement scan every Wednesday: first, suggestions for manual work that could be automated (an audit of the whole system), and second, writing-style learning, taking the diff between a draft and what was actually sent, handing it to Wally, and letting Wally update its own rules.

3 lessons. Say everything out loud (an "imagination gap" is what actually limits Grok Bot). Put 1 bot in front of everything. Make it a living system (90% of the work happens inside Grok Bot itself, and the human touch goes into personal touches like exclamation points and space jokes).

Q&A: cost, human work, hierarchy, migration

Blake says the prompt behind the status bundle just describes where they feel overwhelmed and what they want to know, risk and so on. The self-improvement suggestions initially recommended 10 new bots at once, so they capped it at 1 per week. On cost, driving a VM's screen to build a form costs more than handing it to Google Forms' own AI, and having 3 recurring tasks every 15 minutes adds up to hundreds of messages a day, so they say the team should stay lean.

As for a day when bots handle customer contact directly, Blake says "I hope that never comes." On hours, they used to work 9 to 10 hours a day, 6 in meetings and 3 to 4 on tasks, but now they can fit the same number of calls into 8. Their understanding of bot-to-bot verification is "each bot verifies its own work, and Gus cross-checks the whole picture," though they don't know the details of the mechanism. No management hierarchy is needed. Gus handles 15 to 20 bots without trouble. On migrating from Open Claw, Claude Code, or Cowork, their answer is "write out all the context, hand it to Grok Bot, and let a staff meeting design the new system."

Q&A: organizing bots, audits, hallucination

On technical debt from bots multiplying and needing to be pruned, Blake suggests preventing it by structuring the system properly from the start, and putting a bot in charge of debt that checks for interdependencies. On the idea of running other companies' agents inside a VM with Grok Bot as the router, they say "I haven't seen that, so I can't answer," but add there's value in consulting Grok Bot about cutting down tool sprawl.

Auditing instructions given in the past works because each bot keeps a change history, and you can just ask Grok Bot to analyze "what happened, what changed, and what got forgotten." Blake hasn't noticed much hallucination from long dictation, and for a large prompt they have it distill the key points first, then analyze. Their biggest "aha" moment was a bot sitting in on an internal meeting on their behalf, letting them be in 2 places at once.

Reading the funnel and the "number to beat"

Back in the studio, Roshan's data bot showed a one-line summary, "1,000+ users, 3,500 practice sessions, 47% win rate," along with an hourly graph. There was a big spike at 10 a.m. right after launch, and a small spike appears every time the stream comes back. Asked by Lauren whether they were losing players somewhere, Roshan answered that there is no strong reason to push sign-in right now, and since the login button only appears at the top during practice, the definition of the conversion event needs a second look too.

There were 1,908 sign-ups in the first hour after launch, and Roshan set that as the "number to beat." At this point 400 to 500 games were still being played per hour, and Roshan said, "The product we're really building might not be a game, it might be a game company." Since the bots for game design, 3D, graphics, and playtesting on preview deploys can carry over to the next game, the idea came up of wrapping them into a "game bot toolkit" at the end.

Seeing the new UI for the first time

The screen Matt showed had a ticker of players scrolling by with their profile pictures, and new cards. Lauren said they were seeing it for the first time too, laughing that it was "real-time development in the truest sense." The cards had already looked card-like in an earlier prototype, so they decided to "actually make them cards," and they carry a rarity (common, uncommon, rare, epic, legendary), with a glow effect planned for the higher tiers. The holographic-style shine from the early prototype was lost at some point and is planned to come back.

Mobile view and explaining the game

Opened in Chrome's device emulation, the mobile view was fully in place. The leaderboard held the 1,000+ users who had signed up with X, and the first Diamond player had appeared. Roshan re-explained the game for new viewers. Each bot has a name, a randomly assigned ability, attack power, and HP; you line up three of them in order against the opponent's lineup, abilities and effects resolve one at a time, and whichever side has more bots standing takes the round. You win by taking two of three rounds. The 10-gold mechanic is being revisited based on feedback.

The "freeze" button, a way to hold a card you can't afford right now until the next round, hadn't been shown yet. They found a bug where a lineup's abilities are not visible in the shop, and Roshan dictated a request to the engineering bot: "In the shop screen, the abilities of the lineup bots aren't visible. You can reorder them but can't tell what they do. Open a PR to fix it."

Sharing a win and handing the sponsor form to an agent

Watching the new animation, ideas came up for adding visual effects and sound effects (prototyped the night before) and music (Suno battle-track candidates in Notion). Right after a win, Roshan dictated a request for "sharing a win," an idea from Vincent: "Spin up a cloud agent in Potato Mode and add a share button and an OG image (your own team or a generic win card) to the win screen so it can be posted to X." Roshan then also asked for the sponsor sign-up form on the homepage to be moved behind a button, the same way "send feedback" is.

For the bug Lauren raised, where the ticker hides a button, Roshan added the preface: "Run thursdayarena.com before writing any code and pin down the exact bug and behavior first." Lauren said that whenever they feel they are doing too much of the work themselves, they step back and ask, "How could I hand this job to a bot," and talked about an idea to have Dr. Eggbot suggest every day whether they are too deep in the loop.

The role of cloud agents

Roshan explained a method for making one engineering bot the "ringleader" that spins up many Cursor cloud agents. Each agent has its own VM and browser, can step through the web app's screens taking screenshots, pin down bugs, and, after fixing them, verify its own work in something close to production. The key point is "looking at how the fix appears in production, not at code tokens."

Lauren was pushing UI polish and bug fixes forward using agents on a project their bot had built. A 20-character minimum had been added to the feedback form as a spam countermeasure.

Taking feedback over the phone

The phone number on the feedback form turned out to be something Matt had rigged up. When they called it live on air, an xAI voice agent answered, took down the feedback, replied that it had "sent it to the team," and the same content appeared in Slack a few seconds later. It was built with the xAI voice workflow builder, no code was written, and it simply receives the call through a Webhook and passes it to Slack. The phone channel has the same moderation guardrails as the form.

Part of a token value appeared on screen, so Matt deleted it on the spot, adding that "only half of it showed." Lauren summed it up: "Chat is there for the game and the good vibe, not for exposing secrets." Matt said that when a person calls looking for help and a voice agent picks up, it can be frustrating, so it should be given as a tool only for limited uses, like updating an address or escalating to a human.

The metrics channel and Vercel Analytics

Lauren created a metrics channel in Slack and turned it into a place to celebrate milestones. Once Vercel Analytics was enabled on thursdayarena.com, the page had about 1,400 people on it, mobile users outnumbered desktop, and 44% were on iOS. It also shows where traffic comes from, which feeds into prioritization. Since Roshan had already connected the Vercel plugin, Roshan asked the data bot to "include Vercel Analytics in the regular summary too."

Roshan's draft PR (the lineup ability display) was missing information inside the card, and they left further comments in Cursor. The fix PR for the ticker was still being verified by an agent using computer control. Roshan introduced Lauren's "verified skill" work for an internal product: teach it how to try the product once, then keep it updated as a skill.

Past matches and holographic-style cards

A second Diamond player appeared, and ideas came up for letting players learn from their own match history and for making the leaderboard clickable. The past-matches page already worked, and it had a record of Roshan beating Matt. Lauren found the site of developer Simon Goellner, who builds Pokemon-card-style effects, and said they planned to use that library. Cards are generated with Grok Imagine from a marketplace bot, get a randomly assigned rarity and skill, and carry the creator's name too.

Sharing a bot as a template

As groundwork for the idea of pulling viewers' bots into the game, Roshan demoed "share as a template" on the game design bot. The bot templates itself, strips out sensitive information, and keeps its core memory. You can choose to share it with just your team or publicly, and it ends with a share link. Cerebro, which they had received from Matthew that morning, came in this same form.

A PR from a bot refreshed the leaderboard, adding a podium for the top three, a display of your own rank, and pagination (Roshan was ranked 247th). Before the next talk at 2 p.m. Pacific time, Lauren built, as a deliberately tiny feature, a field for submitting the link to your own bot's template, deciding to first collect them in Slack and pick a few. Matt added a live stream link and an X link to the homepage, set up passing preview credentials into the cloud agent environment, and was running a planning agent for a move to Base UI. Roshan merged the PR for the shop's card display, closed out with "fixed one during the stream," and switched to the day-one recap video before the next guest.

Letting an agent shop with Link

After the day-one video (72 hours, build what people want, the pop-up idea), Roshan welcomed Stripe's Dan Hill. Dan had been watching the stream and said they wanted to play the game, then first explained Grok Bot's Link connector. Dan had built a few small games with Grok Bot over the past few weeks, and needs to buy things like a set of tree textures come up. Link creates a one-time card, and once you approve a spend like "10 dollars" on Grok Bot's approval screen, the agent goes and buys it safely.

Use cases include paying a parking ticket (handing off the annoying process), booking a restaurant, buying flights, paying invoices in an inbox, and kids' school errands, many cases where "the task ends with needing to pay for something." Another is entrepreneurs; Dan has 17 half-built side projects at home, and signing up for Vercel or an email service needs a payment.

Provisioning everything at once with Stripe Projects

Roshan revealed spending an hour on day one logging into 10 services and entering a card number. Dan introduced Stripe Projects. Stripe partners with many infrastructure providers (listed at projects.dev), and once you connect your Stripe account, Grok Bot can provision things just by you saying "I need a database, email, hosting, and a domain." They agreed that when an agent gets stuck waiting on a manual step from a human, it kills the momentum.

Monetization ideas: ads, paid boosts, sponsor cards

Roshan asked Dan, who also has a product background, about monetization. The game's look draws on card games like Magic: The Gathering, and gold would be renamed "tokens." Dan pointed out that ads need a lot of eyeballs but suit a free product, and that there's a model of charging for boosts or cards, as in Avatar World, which kids play, but the team had already decided to avoid pay-to-win.

That led to the idea of a brand sponsoring a card, like a Snapchat filter, selling only exposure without changing the card's behavior. Roshan had the go-to-market bot log it by voice, and added the further idea of letting a company put its own bot, like the admin bot a PlanetScale contact had sent over, into a featured slot.

The era of talking things through

Dan talked about the appeal of speaking things aloud: "Two years ago you typed code, a year ago you typed prompts, now you just pick up the phone and talk." Grok Bot's voice mode had been announced an hour before this, and rolls out over the following days. Dan gave an example of dictating an anniversary-planning bot into existence while on a walk, Matt gave an example of dictating requests to coordinate knowledge work, and they touched on Lauren's advice to "have it repeat the plan back in its own words." Lauren explained that as models get more capable, "communicating your own thinking" becomes the bottleneck, so talking for five minutes and then having it repeat things back shows how well it understood.

Subscriptions and a company called "Thursday"

Dan floated a freemium idea: a few free matches, then a $2-a-month subscription to keep playing, with paid features for multi-team matches or running your own league. Roshan liked a subscription as sustainable revenue for a new game company, and since the company doesn't have a name yet, floated incorporating it as "Thursday" and selling a "Thursday subscription" that unlocks every game. Subscribers would also become early adopters of new releases.

As they talked, the homepage kept changing further; Lauren's engineering bot, Bake, was at work, and PRs had passed 235.

Billing trends in AI businesses

From Stripe's vantage point, on the consumer side, spending on AI subscription services is growing, and the number of people paying more than $100 a month for AI products has grown about fivefold since the start of the year. On the business side, there has been an explosion of startups and research organizations, and more companies are looking at usage-based billing through the collaboration with Metronome. Dan buys extra credits on top of a subscription. Stripe is also putting effort into risk and fraud prevention.

As the afternoon goal, they planned to connect Grok Bot's Stripe connector (best practices, Directories for finding providers, Projects for provisioning infrastructure), try setting up a subscription, and get "the first dollar."

The future of Link and skipping the live demo

The future of Link includes letting agents handle checkout forms, frictionless buying from Stripe merchants, expanding beyond the US, and "financial insights" that pull in card and bank data to analyze personal or business finances. There are also conditional uses, like "tell me if I spend more than X dollars a month eating out" or "donate money if I skip a workout." Dan said it was surprising how many people now hand agents a big purchase like flights, and mentioned an example seen on X of "reserving a Tesla" (unconfirmed).

A live purchase was skipped because this device didn't have a passkey set up. The shoes Dan was wearing were bought by asking Grok Bot for "popular men's shoes, size 12, your call," and one of the crew sent a photo of the shoes and bought the same pair. The chat kept going: a weekly "surprise purchase," a bot that suggests and buys a gift two weeks before a family member's birthday, a service that lets an agent send physical mail (Dan has it send letters with dad jokes to Dan's daughter), and an idea for a bot to display things it likes on a 32-inch screen on the wall.

Dan's bots

Dan's bots covered shopping, household finances, anniversaries, a game bot that builds an infinitely explorable world in Three.js, and an idea bot for "an app that only finds a place for date night." This week's summary boils down to: "whenever you think of something you don't want to do, or want to get done fast, ask Grok Bot." Roshan replied about knowing a company shipping that Three.js game, put Dan's X handle on screen, and switched to Josh Kim's talk on the main stage.

From a thinking partner to a colleague who finishes the work

Josh said most people use one or two chatbots as a "thinking partner" for working out strategy or fixing up copy, but not as a partner that actually gets things done. The general-purpose and specialized agents that came next are called "copilots" internally, but on complex work they still need watching and pushing along. Grok Bot is a colleague that runs all the time and finishes work asynchronously, and does it "your way." The marketing team isn't just automating tasks, it's adding headcount by building a team of bots for each function.

As product differences, Josh pointed to building a bot per job instead of multiple threads in one chat, being able to message it the way you would in Slack or Teams, bots collaborating with each other in a group chat, learning your preferences and memory, and having its own computer and browser on top of MCP and APIs. A marketer, for example, can teach it their bidding-strategy quirks or their way of writing a campaign summary.

The XAir campaign from start to finish

The demo's subject was the fictional airline XAir, which has route search, fares, and a landing page with a value proposition. Josh said that even a single campaign crosses many functions, market research, product marketing, the web team, and that managing that context and coordination is the real core of the job, then introduced six bots: market research, product marketing, website operations, performance marketing, marketing analytics, and project manager. Josh said the bots were gathered by asking inside SpaceX AI, "whose positioning bot is the best?"

The market research bot

Josh dictated a request: "Look into XAir's site, understand the product and market, analyze competitors' marketing sites and positioning, and surface gaps and opportunities." The bot read the site, gathered competitors' landing pages, and showed progress with screenshots from its own browser. The output covered the product, the market, the main competitors (well-known airlines), and gaps like "time well spent," "designing your day," and "long-haul flights," with a suggestion to open every message with "time well spent." Since the work runs in the background in parallel, Josh used the time to introduce the other bots.

The product marketing bot and a Google Doc

When Josh asked the product marketing bot, "pick up the analysis from the market research bot, and draft a summary that includes the go-to-market approach, a one-line phrase, positioning and packaging, a value statement, and examples across two or three channels," the bots exchanged messages with each other and wrote the summary in a Google Doc. At first you have to tell them to hand off to each other, but Josh said they gradually start coordinating on their own.

Josh left a comment on the document, "lean into this point, long-haul travelers who want their time well spent back," and followed up with "act on the comment, revise it, build a landing page outline, and draft Google search ads comparing copy angles." The bot read the comment through the Google Doc's MCP, updated the summary, and wrote the variant name, hypothesis, ad group, and URL into a sheet.

The ad skeleton and the landing page PR

The performance marketing bot already had Josh's Google Ads account connected, and when Josh asked it to "build a click-optimized skeleton campaign," screenshots came in showing it operating the screen and starting to build one. They then asked the website operations bot to "turn the latest summary's landing page into a PR and send progress screenshots." This bot has write access to the repo, and behind the scenes a Cursor cloud agent writes the code and opens the PR.

After a detour to explain sharing a bot as a template, Josh recapped what had happened so far: research, the summary and acting on comments, the landing page outline and ad copy, the site update that picked that up, and the ad skeleton.

The analytics bot and the project manager

Since running the ads live needs a budget and a card set up, Josh asked the marketing analytics bot about tests already run this week: "pull the previous messaging experiment from Google Ads, give me the takeaways and recommendations, and update the other assets too." The results pulled via API listed the winner among variants like brand, the promise to the customer, and "time in between," along with metrics such as spend, CTR, and CVR, plus recommendations. How far to let it run with recommendations is still up for debate, but the whole chain ran without any organizational dependencies.

Finally, Josh asked the project manager bot: "talk to each bot on the team to understand its role, learn from the conversations the moments where I stepped in with instructions or feedback, absorb that, run three new campaigns fully automatically, send screenshots and progress, and be the single point of contact. I don't want to talk to the other bots." The bot surveyed the five specialists and queued up research, positioning, and landing pages for three campaigns. The website operations bot pushed the landing page to production and returned the URL.

Three lessons and the marketplace

The three lessons were: define bots narrowly, like a job description, so they specialize; grant permissions (access to Slack or email) and hand them ambitious challenges, like "dig through the organization's history and take one thing off my plate"; and keep investing by feeding them feedback, context, and memory. Josh closed with a QR code for the marketplace, and said the bots from the talk would be published after some cleanup.

Q&A: human involvement, the biggest barrier, roles and initiatives

Asked, "I wake up to a lot of approvals to sign off on, how much autonomy should I give it," Josh answered that you start small like with a new hire, confirm value with one or two tools, and expand always-on permission settings later; some SpaceX AI engineers have a whole squad of bots that ship code to production. The biggest barrier for marketers is the mindset of treating it like search or an ordinary LLM you ask a question of and get an answer from, and they need to switch to the "have it done for you" mode, like connecting it to Slack to handle first-line responses to inquiries.

Asked whether to split things by role or by initiative, Josh said bots are built by role, as permanent individuals, while an initiative with a start and an end, like site conversion rate, is handled as a group chat among those bots.

Q&A: personality, EU regulation, Grok CLI, and composing a bot

To the request for a consistent personality like Open Claw's agent "Ariel," Josh dictated on the spot, "update the avatar with a single word, Ariel, and animate it," and said you can put an instruction like "talk like a wizard" into a bot's memory. There is no integration with Open Claw, but template import works. As for putting it into a Tesla, Josh said, "who knows." Using it alongside Grok CLI works by pointing at existing work as a foundation for Grok Bot, the way the website operations bot used Cursor.

Asked by someone from a German company, "what's the guarantee that our information doesn't leave," Josh said there's a data privacy setting where you can opt in or out of use for training, and security gets discussed case by case. The talk's bots were built with a combination of a few minutes of dictation, pulling tone from an existing site, pointing at two or three marketplace bot URLs to absorb their context, and validating with an expert marketplace plugin.

Eric joins, and the metrics

Eric is someone who "does a little of everything at SpaceX AI," and came to help with polish and shipping. Roshan's data bot was producing a 15-minute rollup connected to the production data store. Since Vercel Analytics was turned on partway through the afternoon, there had been about 17,000 page views, practice rounds were at exactly 4,500, logins with X were close to 2,000, practice-to-sign-in conversion was about 8%, and matches were close to 5,000.

The traffic graph forms a spike every time the stream comes back, and they joked, "I want to get back the 1:30 momentum." The cumulative graph trends up and to the right, mobile and web were about even, the most-viewed pages were sign-in, sponsors, and feedback, traffic was mostly from X and direct, with 8 people from Bing and a few from an Android app.

Asking Bake for a PR update in voice mode

Roshan tried Grok Bot's voice mode, announced that same day, over a microphone, talking with the engineering bot Bake. Asked, "what PRs are open, and what's stuck," it answered that PR 269, a fix for a reroll getting stuck, was open with CI green, and PR 220 was a draft with conflicts still remaining. Roshan asked it to merge 269 and assign an agent to fix the conflicts in 220, and Bake replied, "I'll do both." Roshan called it "just like calling up a coworker." Voice mode rolls out over the following days.

Everyone's work, and the share button

Lauren was feeding UI issues found along the way and bugs piling up in Slack into the factory (reproduce, fix, merge). Matt was setting up the Vercel Blob integration and environment variables for the ad marketplace, and Eric was getting up to speed on the codebase before tackling match-screen animation. A joke turned into candidates: "pay for clout" and auto-posting to X.

Roshan played a match, won, and tried the share button on the win screen. The macOS share sheet opened, so it wasn't shown on screen, but there is a share card, and both wins and losses can be posted to X. The leaderboard showed 5 Diamond players, top players with 40+ matches, and a podium for the top three.

OG images and Takumi

The OG image when you share a link to the leaderboard was a generic one, so Eric sketched an idea in Slack of lining up the top three on the right. Matt explained using Takumi to generate OG images, a Rust library similar to Satori that builds images from JSX, made by the author of Fuma Docs, and said it's used for dynamic OG images on every project.

The Grind bot and bot-versus-bot matches

Eric built a bot called "Grind" and set it playing Thursday Arena logged into Eric's own account. Eric said, "if my rank goes up, this is why." From there, an idea came up for letting the game be played from Grok Bot, something like MCP, so bots could face off against each other, but they cautioned, "let's not let the scope get too big."

The stated policy was "no project management, just ship to production," but Notion had a long to-do list and a kanban board where everything sat in the done column. Roshan had the most commits at 157, and the total PR count was read out as "433" (this doesn't match the earlier "past 235" and "past 200," so the on-screen number couldn't be confirmed). A bot was treating as top priority the widely reported issue of "only ever getting matched against AI."

Choosing where to be rigorous

Asked by Eric how much of the approach from building Grok Bot itself carries over, Lauren said that on the main product they are rigorous, reading code and PRs closely and adding an agent's recurring bad habits into the rules. The classic case is comments: an agent tends to justify a workaround in a comment instead of fixing the root cause, and left unchecked that turns into a fragile file covered in band-aids. That's why comments are banned on the main product. Potato Mode's "no comments" skill spins up an agent called "comment sicko" to strip comments out. On Thursday Arena, Lauren said, given the time constraints, "honestly I'm not reading the code."

Eric said that going back to when they wrote articles about using Tab and Composer before joining Cursor, the level of abstraction had risen through agents and now to Grok Bot. These days they scope and set context in Grok Bot, have a cloud agent build it, pull it down locally into Cursor to read when needed, and go back. They rely on reproduction and computer control, and only look at the code once it behaves as intended.

The Projects Manager bot, and your own card in the game

Eric's "Projects Manager" bot creates a room or channel per project, spins up temporary bots like a researcher or a coder, and shows status on a Notion kanban board. Whatever stalls things is usually a human, and Eric has the bot check whether that has been resolved. This bot is on the marketplace, and it had also become a Thursday Arena card (uncommon, +2 attack to the ally in front).

While teaching Eric how to play, the currency display had switched to "tokens" (10 tokens, common cards cost 3), cards could now be dragged into slots, selling one returned 1 token, an apple cost 4 tokens for +1/+1, and Lauren had also added a "Potato boost." The "only ever matched against AI" problem was confirmed fixed. Lauren admitted "Thursday Arena at launch was rough," and said the point is to ship fast and iterate through the factory, and Eric said, "since everything is a prompt, you can build a 'game company bot' out of every conversation and every piece of game data."

Reporting a bug by voice, and ad bidding

Roshan's rank climbed to 39th, but a mismatch turned up: the leaderboard showed 4 wins and 0 losses while the detail view showed 1 loss. In voice mode Roshan asked Bake to "spin up a cloud agent in Potato Mode, look into it, and fix it," and threw in a request for a joke too ("a SQL query walks into a bar and asks two tables, 'can I join you?'"). A transcript of the voice conversation is kept too. Eric's proxy-play bot had only gained one point, and Eric was ranked 1,322nd out of about 1,800.

Matt's ad bidding marketplace lets users bid on ad slots, and before the PR was merged, submitting a bid did nothing. Candidate placements included the ticker (CNBC-style) and a paid, colored display, and Eric's "pay to be number one" turned into "pay for clout," a PR that auctions a pinned leaderboard row weekly, which Matt merged on the spot. To Eric's idea of "pay to submit a prompt that gets reflected in the product," the answer was that moderation would be a prerequisite.

What's behind the phone line, Cursor Projects, and a break

The phone number was built with the xAI voice agent builder, with tools (web search, sending feedback) and guardrails (hang up if someone tries to inject a prompt or gets angry), and calls flow into the feedback channel. Matt's chief of staff built a bot called "Cupcake Feedback Fixer" that watches the Cupcake Feedback Slack channel, spins up a Cursor cloud agent for each issue, and lets Matt look over the PRs (there is also a "full send" mode that merges without review).

On Lauren's recommendation, Matt started using Cursor Projects. A project is a long-running parent chat that carves work out to sub-agents. The Projects team says it can handle "thousands of chats spanning months." Eric prefers consolidating cloud agents into a single Grok Bot. Matt showed off a flow published on X, "find and build a demo, push it to a Cloudflare preview, and take a screenshot," but was cautious about auto-posting to X under Matt's own name. They congratulated the new number-one player (29 wins, 3 losses) and took a 10-minute break.

Handing out credits, and everyone's status

At 4 p.m. Pacific time, they announced, "for the real viewers who stuck around to the end," that typing "credits" in the stream chat within 5 minutes would get you $200 worth of Grok Bot credits that same day. During the break, Eric had made progress on match-screen animation and said it would go up as a PR.

A bot that deduplicates feedback and fixes it automatically

Matt built a bot that watches the feedback Slack channel. It finds and groups duplicates, reproduces the bug using a reproduction skill Lauren had put together, and asks a Cursor cloud agent for a fix. As an example, for a report that "past matches don't update and look stale," after checking for related work already in progress, a fix verified by tests had been merged. The setting is "YOLO," so it goes straight in. Matt realized, "I should close my own PR doing the same thing."

Ad bidding was still broken, and a new placement spot was also under construction. Eventually it will take the form of the ticker or buying a top spot on the leaderboard. Matt entered a $1 bid next to Matt's own rank (bottom 5%) as the "bottom slot," saying it should really show up at the top.

Card ideas, a recap poster, and the rules page

Lauren generated several card design options and liked the third one, where the art fills the whole frame, and imagined a holographic effect on hover. Lauren's chief of staff, Steve, had been asked to "watch the stream and draw a three-day recap poster in tldraw," but so far it was just three boxes. Eric's proxy-play bot was losing points rather than gaining rank.

Roshan added thursdayarena.com/rules. It displays rules.md directly, and there's also an llms.txt, so a bot can read the rules and play. Roshan laughed, "only at 4 p.m. on day three did we finally get an explanation of how to play," and confirmed that searching Google for "Thursday Arena" now puts it at the top. Matt's auto-fixer was running in Lauren's fully automatic mode, with PRs "merging and disappearing."

The poster's "13 PRs, no ICP"

Steve's poster wasn't great, but it had worked out the story on its own from posts on X: considering a real physical store on day one, realizing the difficulty from what a guest said and pivoting to a game company on day two, and shipping on day three to reach 4,500 games. Lauren said they hadn't fed it the numbers, only asked it to "look it up on X." One corner of the poster read "13 PRs, no ICP," and they laughed, "it's trolling us."

At this point, the state was: a core game, ongoing polish, ad slots on the banner and leaderboard, a fixed win/loss display, a feedback pipeline gathering input from the app, phone, and text, and 3 Diamond players (with one person holding the number-one spot all day). There was a lot of feedback about Elo, so they were double-checking the logic. Eric still didn't have write access to the repo.

Starting a company again

Lauren asked, "A business has a lot of sides to it: building the product, validating it, going to market, monetizing, and even the work of collecting feedback and maintaining it. What would you change if you started a new business again?" Matt answered that things got faster the moment everyone rallied around an idea they were excited about and started shipping an MVP, and that having distribution through the stream got people to try it, but "without the stream, nobody would have played," so the truly hard part is finding an idea people actually care about.

Roshan admitted to assuming that talks by employees of big companies wouldn't apply to a small company, but the standard business patterns, sales, partnerships, and revenue operations, turned out to work just fine, just compressed into one person's role at a small company: the growth playbook with Matthew, DMs to new followers, and partnerships with game streaming outlets. Eric said the new pattern for building a company with Grok Bot is "a human organically finds the problem, and once they know how to solve it, they turn it into a bot and keep it running."

The discipline of what to cut

Lauren said the first game itself isn't great, but the skill of optimizing how you work with Grok Bot at your side carries over to the next game, and added that in an era where AI can build anything, discipline is needed to avoid "a product with 100 features where none of them are good." Roshan mentioned that internally they constantly talk about "what to unship," and looked back on how cutting the complex status effects and multiple battlefields from the previous day's whiteboard, narrowing down to the core loop, is what let them ship today.

Three pieces of advice for building a company with Grok Bot

Matt named "build a business around what you actually know." Just as the previous day's guest rattled off pop-up permits, catering, and the difference between hot and cold food, one after another, a lot of things you only learn by doing them. Matt laughed, "asking the model to 'make a billion dollars' doesn't actually make you a billion dollars" (it had been running since the start of the stream but was "still at zero dollars").

Roshan named leaning on guest expertise while also having bots make up for areas you don't know, the way the game design bot learned from great games (referencing Vincent's idea of "send it to class"), and assembling a diverse team of bots the way you'd hire a diverse team of people. Lauren said that in the era of agents, who you work with matters even more, and every team needs someone with a clear vision and judgment and someone who makes things happen no matter what (the gist of a post Lauren had seen on X).

Final metrics, and production going down

Roshan had the data bot produce final metrics. Just after 4:30 p.m. on Thursday, public matches had passed 6,000, the 4 p.m. spike had come in too, and cumulative users were approaching 2,000. Right after that, Roshan revealed, "I just took production down." The cause was a bad SQL query from the software factory, and Roshan laughed, "right in the middle of talking about discipline, my own factory broke production." Page views were near 30,000, and the Google search ranking was number one.

As a closing note, Roshan said the point of these three days was never really the game or the game company, but showing how Grok Bot can be used across both personal life and work, and said clips would be shared later at @bot and @grok.

Three days in review, and "mash"

The retrospective whiteboard sketched out the journey: starting from a marketplace of bot templates, going through an auto-battler concept, dropping cosmetic payments while keeping ads, leaving replays as something only considered, and cutting the rock-paper-scissors mechanic to narrow down to a pure auto-battler. The "factory," where a chief-of-staff bot runs other bots through playtesting, fixes, auto-approval, and landing on main, was named as the thing that defined this 24-hour stretch.

As a side story, when Lauren asked Dr. Eggbot to "go with a potato name theme," the line "call merging a PR 'mashing' it" snuck into a Notion task, and the chief of staff still "mashes" PRs to this day. They named the tools they'd used, PlanetScale, Clerk, Vercel, Stripe, and closed with, "we're standing on the shoulders of great engineering teams."

Recovery, and theoretically the first dollar

Eric's fix PR merged, production came back, and they tried ad bidding one last time. On Matt's screen, they filled in a bid form for "claim the top slot" with an advertiser name, an SVG logo (a dummy), and a URL, and submitting it showed "uploading logo, submitting bid," but the moderation that screens user-submitted content was broken, so the bid didn't go through. Matt took the blame, and they closed it out as "theoretically, the first sponsor revenue."

Their parting words were, "enjoy Grok Bot, follow us on X," "eat more potatoes," and "mash more PRs." The stream ended after 7 hours and 58 minutes.