Spawn testing

Part 2 of the fleet essay. The essay is how the fleet ships once it exists. This post is how one Discord bot earned that, and how a weekly watch of new people tells them the product is fun. The numbers are his, except the Valve account, which is Mark Brown’s.

← Spawn fleet  ·  AI Agents  ·  AI  ·  Productivity

How to read it

The post is Jacob (@jsnnsa), 31 Oct 2026, 01:29 UTC. He quotes the 29 Oct essay. About 250 likes, 15 reposts, 230 bookmarks, 12,000 views at capture. He opens with the questions people asked, then “how it’s working and how you can do it too.”

Read it as three streams. The ramp is how the fleet got permission. The scoreboard is the 8-week launch. The watch is the practice he ranks himself: new people, every week, the best thing they do outside the agents. Stop at the questions.

On /spawnfleet, about 60% of code changes open from a production error, a Savi bug, an odd log, or a feature request. That loop stays there. This page is where they decide the product is fun, and where the bill can shrink without anyone leaving.

1. The ramp

His start is small on purpose, and the bill is already named. Bug reports. One Claude in Discord. He reviewed its code for weeks. Autonomy arrived over months. The first month cost about $100,000.

Next to that ramp, the management loop is already hourly. They audit every token against their goals. They double down on what is working. They cut what is failing, fast.

Write this: the one job you would hand a single bot for a month while you still read every change. Name the bill you would accept for that month before you start.

2. The product

In this post he calls Spawn a new social network. The same list says it is “not a tool to make games.” They do not sell anything today. They make money when creators do. The goal is max(fun).

Same morning, asked how Spawn gets to Meta’s size, he answers: be the place you and your friends make and play games. Asked about Ready Player One, he says that is what they are doing, and that a top-down metaverse fails. Only bottom up.

Savi stays the AI inside the product. Creators rate her every 15 messages. The essay’s bar and this post’s bar stay apart. The essay asks for a world a thousand times the size of GTA, on a phone, in one second. This post asks for games 100× the size of GTA 5, loaded in one second, on a five-year-old budget phone, at 60 fps.

3. The scoreboard

Launched eight weeks before this post. No paid marketing. Thousands of daily-active creators, doubling every 10 days. Thirty percent are in Spawn 30 hours a week. Ten percent are in more than 60. The star on those hours is his: recent inputs, and the tab in focus. He says Spawn replaced League, WoW, Roblox, and Minecraft for those people.

Peak concurrent users from about 10 to 2,500. He calls that 250×, and 10 × 250 is 2,500. Average concurrent users from 4 to 650. He calls that 150×. 650 / 4 is about 162×. Tens of thousands of daily active players.

Retention uses different days. At launch the best games had 2% day-1 retention. Those games now have 20–25% day-7 retention and 18–20% day-14 retention. Every week has been better than the last. Line the windows up before you call it one multiple.

Mobile still crashes, on 20% of plays, down from 95% of plays over six weeks. He says real user-made games in a mobile browser are hard. Limits on what you can make would make the crashes easier and the games slop. He is taking the hard path.

4. The watch

Savi scored about 40% in January and 92% the week of this post. January sits months before the eight-week launch he is counting here, so the 40% is from earlier than the concurrent-user chart. He credits the move to watching new people try Spawn every week. He says they stole that from Valve’s playtesting culture, and that it is the best thing they do outside the agents.

Mark Brown, 24 May 2023, is the public account of that culture. Playtesting there means you watch someone play a chunk of the game, then you change the design from what you saw. Gabe Newell calls it the studio’s secret weapon. Kim Swift calls it the most important thing they learned at Valve. Portal was watched nearly every week: test on Friday, talk on Monday, change the game, test again. Ken Birdwell’s rule is to stay quiet while they stumble. The people who made the level are the ones in the room.

Jacob names the culture. He does not cite Brown, and he does not publish the room, the script, or how many new people sit down. The Savi percentage is a separate score, from creators, every 15 messages. The weekly watch is how he says they decided what to change. The two stay side by side.

Write this: one flow a new person will hit this week. The sentence you will write when they get stuck, before you explain it. Who on your team has to be in the room.

5. The bill

Eighteen minutes later, still in the thread, he adds the property. They can cut spend by 95% without firing anyone. The move is to say no to more. If that call is wrong, they can scale it back up 10,000% the next day. He calls that perfectly elastic capacity, and a capability that had not been possible before.

A 95% cut leaves 5% of the spend. Returning that remainder to the old bill is 20×. His 10,000% is the size of the swing he is claiming, written as its own figure. Keep both numbers as his.

Two replies the same morning fill the sentence in. To a spend freakout, he says the bill is payroll for 50 SF engineers, with a large multiple of productivity, and they can scale it up and down hourly. Asked about looser schedules and dynamic incentives, he says they tried that family of ideas and each had issues. If you do not make Claude deploy, he almost never will, because there is always one more fix and one more check worth waiting for.

That deploy line is why the essay calls a quiet hour an outage. The hourly token audit in this post is the other half: every hour, the spend has to still match the goal, or it gets cut.

Write this: one class of agent work you can say no to this week. What the bill does the day you say no, and what you would turn back on the next morning if you were wrong.

What to try, and what to leave

Four things transfer to a team that is not Spawn.

  • One bot, one pile of bug reports, you read the code. Weeks of review. Autonomy over months. He names about $100,000 for month one, so write your own ceiling first.
  • Once a week, watch a person who has not seen it. Stay quiet. Write the moment they get stuck. The maker of that flow is in the room. A rating from people who already use it, on a fixed count of messages, is a second score.
  • Say no, and watch the bill move the same day. The hourly audit is the habit. Elastic means the no is reversible the next morning.
  • Require the deploy. His reply: left alone, Claude waits for one more check. The essay’s hourly ship is this rule with the safety already built.

Leave the rest on /spawnfleet until a broken change can come back in seconds: killed pull requests, a quiet hour treated as an outage, and a token bill at 30× payroll. The weekly watch feeds the taste those judges are there to protect. It does not open the production loop by itself. That loop still starts from an error, a log, or a metric, as on /ryancarson.

Sources

Related on this site

  • Spawn fleet — part 1. Six sittings on the essay this post is answering
  • Ryan Carson — an error that opens the work, beside the weekly watch
  • Company brain — the record a weekly note has to land in
  • GrokBot CEO — a human-shaped roster. Spawn’s ramp starts with one bot and a pile of bugs
  • GrokBot team — the trust curve. His version is weeks of reading the code

Comments

Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login

View comments archive