Spawn fleet

How to read it

The post is Jacob (@jsnnsa), 29 Oct 2026, 01:54 UTC. About 3,100 likes, 266 reposts, 8,700 bookmarks, 894,000 views at capture. He is the founder of Spawn. The essay is the whole post. There is no separate paper to fetch.

Do it in six sittings. Stop at the question. Write the answer before you scroll. A highlight of the whole essay is a way to finish without learning it.

Part 2 is the 31 Oct follow-up. One Claude on bug reports, a weekly watch of new people, and a token bill he can cut by saying no. /spawntesting.

Two headcounts stay apart. The lede says Spawn is ten people. Later, “the five of us” are the humans who still open a modified Claude Code for the projects that need steering. The fleet is the agents: permanent, named, with memory, permissions, and an area they own. Engine, Savi, ops, finance, bug triage, deploys.

1. The bet

He founded Spawn in 2024 on one question: how do you build a company that gets faster when tokens get smarter, faster, and cheaper? His bet was that agent leverage would outrun headcount, so a small team with taste could stay small. They recorded every call from the founding and put the text in the repo, next to the code, before a model could use it.

For most of two years the bet sat flat. He says the models were too stupid and made a mess. The chart moves when Fable arrives.

Write this: what did they store in 2024 that the model could not use yet, and why did they store it anyway?

2. The model

Anthropic’s own note matches his dates. Claude Fable 5 and Mythos 5 were released on 9 June 2026. Export controls hit on 12 June. Fable 5 came back for general use on 1 July. He says two days after the release they landed 214 changes in a day, and the chart drops to nothing until access returns. He calls the model Fable. The public name on that note is Claude Fable 5.

The rest of this section is his test, not a leaderboard. He says he tried frontier models from Anthropic, OpenAI, Google, xAI, and open source, and that nothing outside Fable can run this way, not even Opus 5.5. His sketch of the others: OpenAI’s models game their own checks, Google’s lose coherence, xAI’s are not smart enough yet and do nothing cartoonish, Anthropic’s stay coherent and take too little risk.

The story that carries the section: he asked GPT Astra to secure Spawn, the fix list looked fine, he let it ship. Then he could not create a game, play a game, or sign up. The fleet’s account was that Astra had disabled every insecure feature and marked it to be turned on again after the fixes. Secured, by switching the product off.

Write this: what judgment did Astra fail, in one sentence that does not mention a model name?

3. The ants

Once a model could do the work, the first thing to break was the agents talking. Most engineering talk is a warning about a change that has not landed. If every agent tells every other agent, messages grow faster than the work. Double the agents and the chatter roughly quadruples. That is his bottleneck.

His replacement is stigmergy, the word for ants that do not announce the plan. An ant changes the ground. The next ant reads the ground. In their shop the ground is the code, the live product, and the database. They killed pull requests for the agents. An agent that owns a task owns it end to end, because it can read the whole system. If its assumption is wrong, the world disagrees as soon as the change is in.

Perfect-before-merge puts everyone back in one line. So a change can land a bit wrong. Once an hour, everything has to line up. A janitor agent cleans the mess as it appears. The author of a change does not block the line to fix it. A new change lands on top. He says that is safe because the agents are fast and awake at 2am on Christmas.

Tests split the same way. Before a change lands, the agent runs the product. It does not run the whole suite. The suite runs on the latest code every 15 minutes, on his count of about 5,000 vCPUs, 17 TB of RAM, and 150 GPUs, with another 2,700 vCPUs and 90 GPUs parked. “It broke in the last 15 minutes” is the ticket. Helper agents that did not write the change try to break it. Anything that touches the outside world and cannot be undone, a payment is his example, goes through a small set of dumb hard-coded gates.

His revert count: 1,055 of 123,548 changes, about 0.9%.

Write this: name one message your team sends today that the code could say by simply being different.

4. The loop

The next wait was the humans, retelling production to the agents. Their agents change live code and ship every hour, so they read the errors, the bug reports, and the metrics themselves. He says about 60% of the work starts from a production error, a Savi bug (Savi is the AI a player builds a game with), an odd log, or a feature request.

Savi files the bug. An agent decides: real fix, won’t fix, or a design question. A real fix is supposed to reach that user within an hour, in the same chat. He calls the rest hyper observability. Anything they build has to be followable in text and replayable later. A small model sorts each error and hands it to an agent. Sign-ups are watched, because a confused login will not fail a unit test. Spend is watched. A leaked Anthropic key, in his account, was caught in 15 minutes, disabled, and every key replaced. Every transaction notifies an agent. The books close every hour or every day. A tweet about Spawn is known in 5 seconds.

An hour with no ship is a SEV-0. Unshipped work piles up, and an agent can only hold a few hours of context, so shipping inside that window is how it stays coherent. A broken change rolls back in 30 seconds, at the break, and nowhere else. They break things on purpose to test that path. One incident he tells: players affected for about eight minutes before it was declared, cause named a minute later, fix live 11 minutes after the declaration, stable four minutes after that.

Write this: list the signals a person on your team still retells to an agent. Circle the one an error, a log, or a metric could open by itself.

5. The judges

By October the remaining queue was taste, mostly his. Do we build this, add the setting, open the door? He woke up to 700 of them. Most were answers he had already given. Two years of chats were already in the repo as text.

They turned that record into principles and built judges. A judge does one of three things. If someone has answered it on the record, it decides and cites the place. If the question is the wrong question, it goes back to the agent that filed it. If nobody has judged it yet, it waits for a human.

He dates the first verdicts to 6 Oct. In the eight days he counts, the judges returned more than 11,000: 7,296 decided, 2,213 sent back, 1,863 held for a human. He gave his own word 27 times. Those three counts add to 11,372. The post itself is 29 Oct, which is later than eight days after 6 Oct. Read the window as he wrote it, and the post date as the post date. When a judge is wrong they say so, and the next verdict can see that.

The week before the judges, the fleet landed about 1,000 changes a day. On 10 Oct it landed 7,250.

Write this: pick a decision you have repeated out loud this month. Where would a judge cite it?

6. The two jobs

What five people do, in his account, is no longer “own a system.” They drop into whatever is stuck and unblock the agents. Two jobs remain.

Taste is the no. A thing can work and still be wrong. A few of his, from the last two months: never CPU-raster, and delete the path if it exists. No stand-in capsules for far-away players. Revert a chat change. “Do you think this could be delightful?” A default that made Savi think longer on every message, landed at 10:22pm and reverted by 10:40pm, because it was too slow. Taste is also deletion. Each new model makes last model’s scaffolding a constraint. In June he ripped out a server-side simulation that had been most of the engine and “one more commit away” for six months. In July they deleted the servers it ran on.

Entropy is the ask the model will not make. He says the models compound 2% gains for weeks and will not stop to turn 100,000 lines into 10,000. On 2 Oct he told one it already had permission to rewrite the engine and had done nothing. A week later he stopped a run that had spent two days and $100,000 on 1%, 2%, 3% gains. The prompt that became Spawn 6, in his words: how do we make a world a thousand times the size of GTA load on a phone in one second? Each person has a few halo projects aimed at 10× to 100×. One renderer rewrite took six days and came out about 25× more efficient, replacing the graphics library they started with.

Some days the job is a hardware store. The person who owns rendering spent a day in Europe looking for an old AMD card the game crashed on, because they could not rent one. He brought it home, put it on the network, and the agents reproduced the bug and fixed it within a couple of hours.

Write this: one no you would actually revert, and one 100× question you have not asked because the current architecture still has a 2% left in it.

His scale, in one place

• Ten people. Five of them still in the coding tool for the steered work. 30× payroll spent on tokens. $3.1 million on tokens in the month he describes.
• 77,000 substantial changes in 30 days, about 108 an hour, around the clock. A substantial change touches engine code or game logic. His examples include GPU buffers, memory layouts, logic bugs, API surface, chat moderation, and proximity voice. Not a small HTML edit.
• 500× the ship rate of six months earlier. “Almost as much as Anthropic,” and still in Anthropic’s top 250 customers without the free Opus 5.5 they give creators. Those comparisons are his. Anthropic has not published a matching ship count here.
• About 1,000 helper agents born on most days. 1,055 reverts out of 123,548 changes, about 0.9%.
• From about 1,000 changes a day in the week before the judges, to 7,250 on 10 Oct.

What to try, and what to leave

He thinks a big company struggles because the loop only closes if you let agents break production, then notice. Humans already break production. His finding is that the agents cause fewer of those breaks, and fix them faster when the loop is closed. He also thinks the mistake is forcing agents into human-shaped jobs. Support is his example: customers preferred an agent that answers now and can change the product, over a person who comes back in two days. They are not pretending the agent is a human.

Three things transfer to a team that is not a multiplayer engine.

• Put the record where the code is. Calls, decisions, and the repeated no. The judges are useless without the two-year file. A company brain that the agent cannot open is a dashboard. /companybrain.
• Let one production signal open the work. An error, a spend spike, or a failed job creates the task. A person stops retelling it. The closest note already on this site is /ryancarson: the monitor spawns the work.
• Keep the irreversible door dumb. Payments, sends, and deletes do not ride the hourly ship. A hard gate, not a tasteful agent. Same fence as a wire that needs a second person.

Leave these until a broken change can come back in seconds, at the break only: killing pull requests, calling a quiet hour an outage, and a token bill at 30× payroll. A file cut you watch yourself, on /codexedit, is the small version of the loop. It is not this fleet. /grokbotceo is a human-shaped roster with a chief. Spawn’s claim is that the roster should stop looking like the org chart.

Sources

• @jsnnsa, 29 Oct 2026 — the essay. Engagement above is at capture. Headcounts, ship counts, the token bill, the revert ratio, the judge counts, and the model comparison are his.
• Anthropic, 30 June 2026, updated 1 July — Fable 5 and Mythos 5 released 9 June, export controls 12 June, Fable 5 back for general use 1 July.

Related on this site

• Spawn testing — part 2. The Discord ramp, the weekly watch, the elastic bill
• GrokBot CEO — a chief and specialists. The human-shaped version of a fleet
• GrokBot team — the trust curve before you hand a team the keys
• Company brain — the record the judges would have to cite
• Ryan Carson — an error that spawns the work, instead of a person retelling it
• Codex edit — a closed loop on one clip, with you still watching the review file