A 10-person gaming startup called Spawn spends 30 times more on tokens than payroll and has an agent fleet shipping 77,000 code changes a month. Its founder's two-year playbook is a field guide to what AI-native actually means, and why most companies will find it hard to copy. In the headlines: an AI startup bids for a bankrupt white-shoe law firm, and tech earnings expose a widening rift between hyperscalers and investors.
Truly AI-native means tearing out the systems built for humans, not speeding them up.
Most companies doing "AI transformation" use AI to do the same things faster, with an org chart someone from 20 years ago would still recognize. Spawn went the other way. It runs a persistent fleet of agents with names, memory and ownership of real functions. It killed pull requests, closed the feedback loop so player agents report bugs straight to the fleet, and turned two years of founder decisions into precedent that AI judges apply. Almost every breakthrough came from removing scaffolding designed for human work. What remains for humans is taste, deleting what new models make redundant, and supplying the ambition agents won't come up with on their own.
Most companies, AIDB included by the host's own admission, use AI to do the same things faster and better. Outputs stay similar, and the org chart would be recognizable to someone from 20 years ago. Spawn is what it looks like to drop the old startup rules entirely.
The 10-person game-creation startup runs a fleet of agents that persist instead of being spun up and torn down. Each has a name, an identity, memory and ownership of a function: game engine, bug hunting, even finance. The agents get Sundays off.
That spend makes Spawn one of Anthropic's biggest customers. The fleet shipped 77,000 code changes last month, about 108 an hour. Founder Jacob Sanbury once assumed he was using agents wrong because his bills dwarfed what friends inside big tech were seeing.
Fable 500×'d the changes Spawn could ship, but the company had to rebuild operations around it. Sanbury's comparison: OpenAI's models game their own checks, Google's don't stay coherent long, and xAI's aren't smart enough yet. Anthropic's stay coherent and do the work, though he says their biggest weakness is not taking enough risk.
— Jacob Sanbury, founder of Spawn. A researcher at Anthropic told Sanbury that of all the customers they talk to, Spawn is the one that works most like Anthropic. Each day the fleet spawns thousands of helper agents for granular work, while the humans use a modified Claude Code for bigger-picture steering.
Agents spent most of their time telling each other what they were changing, and doubling the agents roughly quadrupled the messages. Unlike specialized human devs, agents understand the whole system. So Spawn removed pull requests and lets agents work directly in the codebase, spawning a fixer whenever something breaks.
Test suites run every 15 minutes and the fleet has to have everything in line once an hour. The approach turns out to be surprisingly resilient, with no humans paged at 2 a.m. The cost is heavy: 5,000 CPUs, 17 TB of RAM and 150 GPUs on AWS, plus 60% surge capacity on standby.
Instead of humans relaying production faults to agents, players' in-game agents detect bugs and file reports automatically. A small model triages each report or feature request and routes it to the right fleet agent. The product is designed to be operated and observed by agents, which makes it close to self-healing.
Because agents have limited memory and context, the system is more stable when it moves steadily than when it freezes and then lurches. In one typical incident, the issue was detected eight minutes after it hit a player, the cause was named a minute later, and a fix was live eleven minutes after that.
The agents were finding questions faster than Sanbury could answer them, so Spawn formalized two years of decision logs into principles applied by AI judges. In the first eight days the judges made 11,000+ decisions: about 7,300 from precedent, 2,200 sent back to agents, 1,800 escalated to humans. Sanbury decided only 27, and daily shipped changes rose from about 1,000 to over 7,000.
Humans drop into whatever situation needs them, get up to speed fast and unblock the agents. A big part of the job is taste, which includes telling Claude no even when something works. Another is deleting: in June, Sanbury scrapped six months of work covering 70% of the engine because a new model had made it redundant.
Agents are bad at noticing they've been heading the wrong way for days, and they won't propose radical ambition on their own. The prompt behind Spawn's best build yet came from the humans: how do you make a world a thousand times the size of GTA load on a phone in one second?
Sanbury argues none of this works unless you close the loop, and that takes risk appetite. Humans break production all the time, since changes are the top cause of outages. In Spawn's experience, agents break things less often and fix them fast, as long as the loop lets them notice. With the right structure, every new model makes the whole company significantly better.
Spawn's agents handle customer support. They are sometimes a little wrong, but they answer right away and can act on the spot. When Spawn has taken them off support, customers got upset: they'd rather have an instant fix than a human who comes back two days later. Spawn doesn't pass its agents off as human.
First, a system like this is hard enough to build from scratch. If the host were replicating it, he would start a new division or product rather than retrofit an old one. Second, Spawn's self-healing depends on users' agents reporting bugs in real time, a feedback loop most products don't naturally have. The open question is whether you can design one in.
1) Give agents ongoing responsibility instead of one-off tasks. 2) Let the work carry the context through a shared, current record. 3) Make results observable. 4) Turn decisions into precedent, so each one also builds the canon. 5) Keep human attention on direction, because ambition and long-horizon thinking don't come naturally to agents.
Weil, Gotshal & Manges, the 95-year-old firm that handled the Lehman and Enron bankruptcies, filed for Chapter 11. The cause was departing partners and a failed merger, not AI. Legal AI startup Crosby, last valued at $400 million with about 50 lawyers, has submitted a letter of intent for Weil's trademark, brand and internal data, and hopes to hire some of its 1,200 lawyers.
Firms are thought to have almost no equity value once the rainmakers leave. But precedent documents, contract templates and workflows for elite bankruptcy work could be worth tens of millions of dollars to the right AI buyer. Lawyer Harneet K. pointed out that Google recently bid $10 million for Spirit Airlines' far less specialized operational data.
Gen Legal CEO Ryan Walker says his company looked at the deal and passed. His argument: the data belongs to the clients, the talent has already left, and it is better to recruit Weil attorneys directly. Crosby's investors at Lux and Index, meanwhile, are calling the bid a "master stroke."
AWS grew 39%, its best in almost five years, but Amazon's overall revenue and net income missed forecasts and the stock fell 6% after hours. Intel's data center business grew 78%, yet a soft forecast sent shares down 23%. Google dropped 12% in the day session, losing half a trillion dollars in market cap.
Amazon is $42.4 billion cash-flow negative over twelve months. Intel's foundry is growing 58% while posting a $1.9 billion quarterly operating loss. The AI numbers in the filings don't point to a bubble, but if tech keeps sliding, expect the financial press to call it one.
— Andy Jassy, Amazon CEO, on the earnings call. The hyperscalers are not backing off. Analyst Andreas Steno Larsen calls it one of the fiercest battles he can recall between capital-market investors and tech executives who keep spending despite the market's objections.
A Time profile reported that Trump asked Grok about the implications of forcibly removing Maduro. Reps. Raskin and Garcia, who could chair Judiciary and Oversight after the midterms, say chatbot logs are U.S. government property. They call relying on a "highly manipulatable and demonstrably error-prone" technology deeply troubling. The White House says it preserves all records the law requires.
The Times reports that an executive at Singapore-based Megaspeed, under U.S. surveillance on suspicion of diverting chips and passing information to Beijing, had a close relationship with Huang. Huang reportedly never received a defensive briefing. Nvidia says Megaspeed received only ordinary business information and that the diversion claims were debunked. The host's caution: the reporting is necessarily thin, but it comes from the Times, not a random poster.
