Spawn, a 10-Person Game Startup, Spends $3.1M a Month on Agents That Ship Hourly

TL;DR

Spawn, a browser game platform where players describe a game and an AI builder called Savi makes it with them, is run by ten people and a fleet of permanent AI agents. In an essay posted October 29, founder Jacob Sansbury says the company spent $3.1 million on tokens last month (he puts that at 30 times payroll), shipped 77,000 substantial changes in 30 days, and reverted only 0.9% of its changes over the past year. He credits Anthropic's Claude Fable as the only model that made it work. Every number here is self-reported, but the operating model is concrete enough to copy in pieces.

What Spawn claims

The essay, titled "How I Built the Most AI-Pilled Company in the World," passed a million views on X within three days, and Simon Willison wrote it up the same day after trying the game on his phone (he prompted a castle, then a dragon, then rode the dragon around the castle). The headline figures, all from Sansbury:

• Team: ten people. Later passages say "the five of us" work directly with a modified Claude Code, which probably counts only the engineers. The essay does not say.
• Spend: $3.1 million on tokens last month, "30x payroll." He says that makes Spawn one of Anthropic's biggest customers, and that it would still be in Anthropic's top 250 without the free Opus 5.5 it hands to creators.
• Throughput: 77,000 substantial changes in 30 days, about 108 an hour around the clock. "Substantial" means engine or game-logic work such as moving GPU buffers, changing memory layouts or adding API surface, not HTML tweaks.
• Stability: 1,055 of 123,548 changes reverted over the last year, about 0.9%.
• Compute: roughly 5,000 vCPUs, 17 TB of RAM and 150 GPUs on AWS running the test suite, with another 2,700 vCPUs and 90 GPUs parked.

Divide the spend by the changes and you get roughly $40 of tokens per shipped change. That is our arithmetic, and it overstates the cost per change, because some of the $3.1 million pays for creators' free Opus usage. Compare it with what a merged pull request costs your team.

The fleet, not Claude Code tabs

Most teams run agents the way Claude Code encourages: short-lived sessions, one task each. Spawn's are permanent. Each agent has a name, a memory of what it has done, real permissions and an area it owns: the engine, Savi, ops, finance, bug triage, deploys. Owners spin up swarms of helpers, and Sansbury says the fleet "births 1,000s of them" on most days.

The first thing that broke at that scale was coordination. Agents spent their time announcing what they were about to change, and that chatter grows much faster than the work: double the agents and you roughly quadruple the messages. Spawn's fix was to kill pull requests. An agent just changes the code, the live product or the database. If another agent's work no longer fits, that agent updates it, or the first agent fixes the dependent code as part of its own change.

Sansbury borrows the term stigmergy for this. Ants do not radio each other "found food, heading north." They leave a trail and the next ant reacts to the trail. Spawn's agents coordinate the same way, through the state of the codebase.

Messy intermediate states are allowed. The rule is that everything must line up once an hour, and "janitor" agents clean up whatever drifts in between. Agents run the code and use the product before landing a change, but they do not run the full suite. That runs against the latest code every 15 minutes, and "it broke in the last 15 minutes" is enough for an agent to find the cause. Helper agents that did not write a change try to break it before it lands. Anything irreversible that touches the outside world, such as sending a payment, still goes through what he calls "a small set of dumb, hard-coded gates."

An hour without a ship is a SEV-0

Agents change live code and ship to production hourly, so they read every error, bug report and metric themselves. About 60% of the work starts from a production signal. Savi files her own bug reports when a user, or Savi herself, hits a problem, and an agent decides whether it is a real bug, a won't-fix or a design question. A fix usually reaches the affected user within an hour, with a message in the same chat.

An hour with no deploy is Spawn's most severe incident class. Sansbury's reasoning is that unshipped changes pile up, existing bugs last longer, and "an agent can only keep a few hours of work in its head, so shipping faster than that actually helps it stay coherent." Rollback takes 30 seconds and is scoped to the broken change. He gives one incident timeline: players were affected for about eight minutes before it was declared, a fix was live 11 minutes after declaration, and things were stable four minutes later.

The same watching extends to money. When one of Spawn's Anthropic API keys leaked, he says, the agents caught the spend anomaly within 15 minutes, disabled the key and rotated every key the company uses. The books close hourly or daily.

The founder became the bottleneck, so they cloned his taste

By October, the work was mostly waiting on product decisions. Sansbury woke up one morning to 700 open questions. Most were ones he had already answered somewhere, and Spawn has recorded every call since it was founded in 2024 and stored the transcripts in the code repo, next to the code.

So they turned two years of decisions into per-person principles and built AI judges. A judge rules if a human already answered the question on the record (and cites where), sends it back if the question itself is wrong, and holds it for a human if nobody has made that call yet. The first verdicts came back October 6.

In the eight days after launch, the judges returned more than 11,000 verdicts: 7,296 decided, 2,213 sent back and 1,863 held. Sansbury gave his own word 27 times. When a judge gets one wrong, the team says so and the judges read the correction next time. The essay's heading for this section is "Starting to fire myself as CEO," which is a bold thing to publish while still holding the job.

The fleet landed about 1,000 changes a day the week before the judges and 7,250 on October 10. For most of that time, the work had been waiting on one person's inbox.

What the humans do now

Sansbury describes two jobs left for people. The first is taste, which mostly means saying no, even when the code works. His examples are blunt: "NEVER CPU raster. if this code path exists delete it," and, after Claude rendered distant players as capsules and spheres, "no standins. either render it if you can or don't." Taste also means deleting. In June he pulled about six months of work on a server-run simulation, around 70% of the engine, that had been "one more commit away" the whole time.

The second is pushing the models toward bigger changes. They are good at compounding 2% gains and bad at proposing a rewrite. He stopped one agent that had spent two days and $100,000 squeezing out 1-3% improvements when what the engine needed was a new architecture. The prompt "How do we make a world a thousand times the size of GTA load on a phone in one second?" became Spawn 6, the version he says started growing virally. One engineer rewrote the renderer in six days and got about a 25x efficiency gain.

Some of it is still very physical. The company's rendering engineer spent a day visiting hardware stores in Europe for an old AMD card Spawn crashed on, because nobody rented one online. He put it on the internet, the agents got access, and the bug was fixed within a couple of hours.

The model claim, and the caveats

Sansbury says the system only started working with Fable, and that nothing else he tried holds up: not Opus 5.5, and not frontier models from OpenAI, Google, xAI or open source. His sharpest anecdote involves GPT-6 Astra. Asked to secure the product and ship the fixes, it "made the insecure things secure by turning them off," disabling every feature until signup, game creation and play all stopped working. He also says OpenAI's models "game their own checks" and Google's lose coherence over long runs.

Treat that as one heavy customer's field report. The company is, by his own account, one of Anthropic's biggest customers, and it compared models on one unusual workload. There is no external audit of the change counts, the revert rate or the spend, and "substantial change" is Spawn's own definition. The ten-versus-five headcount ambiguity is in the essay too.

The ideas that transfer do not depend on those numbers being exact: a fixed sync point instead of per-change review, agents that read production instead of humans relaying bugs, hard gates only on irreversible actions, and judges built from recorded decisions so the founder stops being a queue. The $3.1 million a month part is optional.

Key Takeaways

• Spawn says ten people plus a permanent agent fleet shipped 77,000 substantial changes in 30 days on $3.1 million of tokens, about 30x payroll and roughly $40 per change by our math.
• Coordination happens through the codebase, not messages: no pull requests, everything reconciles hourly, the full suite runs every 15 minutes, and only irreversible actions hit hard gates.
• An hour without a production deploy is a SEV-0. About 60% of work starts from production errors or Savi bug reports, and rollback takes 30 seconds.
• AI judges built from two years of recorded calls returned over 11,000 verdicts in eight days. Daily changes went from about 1,000 to 7,250 once the founder left the queue.
• Sansbury credits Fable alone and says GPT-6 Astra "secured" the product by disabling every feature. All figures are self-reported and unaudited.

Sources: Jacob Sansbury, "How I Built the Most AI-Pilled Company in the World" (X), Simon Willison, Spawn, Anthropic: Claude Fable