10 People, $3.2 Million a Month on Tokens, and They Deleted the PR

10 people, $3.2 million in monthly token spend. Anywhere else, that number would look like out-of-control costs; at Spawn, it is day-to-day operating expense.

Spawn is a platform for creating large-scale multiplayer online games and one of Anthropic's major customers. Jacob wrote a long post reviewing how they built this system. Zhang Lixing shared it and called it the article of the year. The reason is simple: outside OpenAI and Anthropic, it offers an intuitive demonstration of the AI-native organizational form. Small team, lots of Agents, Agents First, ample context and loops.

Spawn had context awareness from day one. All call recordings and code are archived—not for immediate use, but to feed to agents once their capabilities are strong enough. The hard part is not technology; it is that most teams will not accumulate data in advance for a model that does not yet exist.

On model division of labor, Jacob's assessment of each company's models is direct: astra is overly cautious, grok is not smart enough, gemini is poor at long-horizon tasks, and the anthropic series lacks ambition. This long-running fleet of agents would only really be able to run once fable5 came out. In his words, Rich Sutton's "The Bitter Lesson" has proven true again: when the model is strong enough, it crushes all engineering patches.

The most interesting move is deleting PRs. Jacob removed PRs from the process because of the communication bottleneck between agents. Humans need PRs because, after specialization and division of labor, different people own different modules and must align through review. Agents are different: they have deep understanding of the entire system, so communication has no practical meaning. Making them review each other is making way for a bottleneck that should not exist in the first place.

With PRs deleted, stability relies on three other things: full test runs every 15 minutes, ample rollback capability, and tiered controls, with sensitive operations such as payments handled separately. Over the past year, there were 120,000 changes and only 0.9% were rolled back.

Even more extreme is the self-healing software system. In the traditional process, agents run on test machines and wait for human feedback on bugs, making humans the bottleneck. Spawn's process is: an issue affecting players is detected after 8 minutes, named and categorized within 1 minute, then fixed 11 minutes later, after which the platform runs stably for 4 hours.

Other parts of the process are so fast that problems are generated faster than the CEO can make decisions. When Jacob wakes up in the morning, seven hundred issues are waiting for him to answer. The solution is to train a judge agent from historical decision records to filter out which decisions really need a human to decide and which can be answered directly based on experience. The number of cases requiring human handling dropped by an order of magnitude.

The human role is redefined as two things: taste and ambition. Engineering capability is becoming cheap, and demand is flourishing; the hard part is knowing when to say no. Ambition means not going for a 2% improvement, but for a 100x leap; how to do it specifically is Claude's job. Claude engineering lead Felix Rieseberg wrote "How to be more ambitious," breaking ambition into four dimensions—speed, height, depth, and breadth—which aligns with what Jacob discussed.

Zhang Lixing wrote an article a couple of days ago about how enterprises can use Context and Loop to let their organizations benefit from stronger models. He used a table to compare his approach with Jacob's:

His judgment is that the physical world has too many constraints, so what he can implement is a second-rate solution, while the software industry can directly pursue a first-rate solution. So when he saw the real Spawn case, he was very happy. Later he added four more points: treat the organization as an evolvable system; model division of labor must combine scenario and cost—for high-value scenarios, be willing to use the strongest model, and the overall cost is actually the lowest; beyond context and loop there is also autonomy—while delegating power, how do you make the organization better and better without sliding into chaos; agents are responsible for doing things right, while humans decide what the right things are.

Jacob says this system is hard for large companies to copy. To copy it, three conditions must be met: you must build Loops, and the important work for humans is to remove bottlenecks in the Agent loop; you must be willing to take risks, dare to let agents into production environments, and still be able to contain the blast radius; and you must design the organizational structure based on agents—directly transplanting the human organizational structure will not work.

10 people burning $3.2 million in tokens per month—that number itself is not the threshold. The threshold is whether you dare to delete PRs, whether you are willing to let agents touch production environments, and whether you can accept changing from an executor into the person who says no.