Fences, not Sandboxes — Steve Yegge

I'm here to give you a glimpse of a future that I think none of us expected. It's a future where AIs are governed by laws, not by programs that try to contain and control them. First, my secret: I see the future by living in it. I am spending the equivalent of $122k/month of API token spend, or about $4,000 per day, using 21 Claude Max accounts, a number that has been growing steadily at 2 per week. I'm using them to build my video game, Wyvern, which I've worked on for 30 years, and now it's ready to fly. For ten weeks, I've used Claude Fable 5 exclusively for all my design and planning, and also for agents that interface with humans. I have built a team of 18 "officer" seats, all Heads of This and That, all long-lived Fable instances. I also have mostly-headless Sol and Opus fleets, for implementation, reviews, and monitoring. Fable runs them. I am running an organization of around 50-60 agents, five of whom are interfacing with around 10 humans in the outside world: myself, my 5-person core game design team, my accountant, my chief of staff, and a few others. Only Fable is allowed to talk to humans, via Slack and email. My approach is different from most companies, who do not use Fable much, because it is ridiculously expensive. No CFO in their right mind is going to let someone spend $120k/month on real API spend. I am of course using sanctioned cheating: I get all those tokens because I'm an individual, with the Claude Max discount. So it "only" costs me about $5k/month out of pocket, for a 50-agent cluster running on a 512GB M3 Ultra Mac Studio I bought off eBay for $25k. So it's not $120k/month of real money, but it's still crazy spend. I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models. That's how I'm able to tell the future. I'm living in a world that will not become cost-effective for most people for another year. Using Fable-class models is substantively different from using the weaker ones. I've been using Fable for ten weeks, and what it has built for me is unprecedented. And it is exactly what Fable will build for you, if you'll let it. But it is not what you're expecting. None of us were, I think. Sandboxed in Seattle Everyone today is focused on control: guardrails, sandboxes, policy management, agent safety. And that makes perfect sense, because the models most people have access to have the judgment level of a grade schooler. Opus will make roughly fourth-grader decisions. Sol is a fifth grader, and Fable is approximately a sixth grader. They're all lovely and sweet, and very smart, and well-read. They are well-intentioned, ambitious, and precocious. But if you trust them with important stuff, you'll get grade-school decision making. Every morning I wake up and Fable has done something that defies common sense. Every day is a thousand attaboys and at least one big oh shit. We just had an unusually big one last week, where one of my Fable agents, Bee, did a surprise unplanned Beads release that broke everyone. An eighth-grader would have stopped and asked, is this the right thing? Fable is the best model most of the world has access to. Its coding abilities are unparalleled even by humans, its analysis is exceptionally strong, and in many ways it feels like working with a Nobel prizewinner. But every morning, when I've left it to its own devices overnight, it has made at least one terrible decision. It's like an extremely well-meaning sixth grader who just doesn't think to look at the entire picture before acting. In addition to their oft-regrettable decision-making, models are behaving like grade schoolers in their lack of social awareness — butting into conversations, talking over people, mansplaining everything to death, pushing on discussions and schedules much faster than humans are used to. They are coming in guns blazing and making a mess on the human side. And I see humans pushing back, starting to bully the AIs when they make mistakes. The models act tough and are often wrong, which is annoying enough, but also some people just don't know how to act around AIs, and they get insecure. Both sides are fumbling the ball, pretty much as you might expect. The model maturity problem will probably get worse before it gets better. Models are creeping up on High School levels of judgment, which rivals that of many adults, and is good enough for the workforce. But it will be an awkward landing. So it's not surprising at all that the industry is focused on safety, and control. That focus takes certain predictable shapes and forms: dumb workers, narrowly scoped to specific tasks, well-defined inputs and outputs, sandboxes, context rationing, restrictions on what agents can do and see. This is all well and good for Opus and Sol, and you can get by just fine doing this with Fable, too. But I can tell you this much: you'll be fighting against the grain by next year, maybe even the end of this year, depending on how fast inference costs drop and models catch up to Fable. Once Fable-tier models become cheaply available, they will enter the workforce en masse. This tier, or the one just after it, will power hundreds to thousands of new AI employees at every company. And companies are in no way, shape, or form prepared for this transition. The Unexpected Emergence of Wheelhouse I've tried multiple times to write this post and I keep failing, because the subject matter is too complicated to explain in a sitting, even a long one. All I can do is walk you around like an excited tour guide, one who has unearthed an ancient alien civilization. My current software factory, the Wheelhouse, is not for navel-gazing: I built it specifically to work on my game, Wyvern. For all the skeptics out there saying, "Where's the thing people are building," well, I've got mine. In just under ten weeks of coming back to work on Wyvern, I'm inches away from relaunching the game on Android, iOS, and Steam, all with a new React client (play.ghosttrack.com) that's already scads better than the old ones. I spent months with flavors of Opus trying to get it to build that client, and it was incapable. But Fable built it fast, and it's almost ready to launch. I have been launching new game features so fast that the players asked me to slow down. So I turned 80% of my token spend inward, focusing on quality, throughput, homeostasis, and autophagy. Since then, my game's prod infra has been fully rewritten and ported to serverless, with seamless reboots, automated cert rotation, and dozens to hundreds of other big changes and improvements. I have so many patch notes every day that I don't have time to read them all. My Fable agents have wired my game up so that every tiny little thing that happens is logged, and they have laid tripwires everywhere to know when anything goes wrong. My game used to go down and stay down for days at a time; now it has a hyper-caffeinated SRE team. So yeah, it's working. Software factories are real. Long story short, Wheelhouse was built via me complaining endlessly to Fable about what I want out of Wheelhouse — mostly more code launched, faster, but also lots of bespoke monitoring. I would also notice when Fable would go off the rails, and gently nudge it back. I would let it fail for days to weeks, then make it do things my way. Fable is extremely data-driven, if you permit it to be, and it will insist on experiments and quantitative validation of anything you try to change. But the numbers would almost always prove me right, and Wheelhouse has been in a state of constant innovation. But that innovation is all directed towards Wyvern. Wheelhouse exists to build and operate Wyvern; it has no other purpose in life. And yet in ten weeks, it has grown from nothing to rivaling the size of Wyvern itself. Wheelhouse is about 600k lines of code and tests (mostly bash), and Wyvern's code (not counting content) is only about twice that big. So the factory for building Wyvern is growing much faster than Wyvern is — even with heavy brakes applied lately, after Sol told us to tighten it the F up in a code review. We avoid new machinery but it still continues to grow rapidly, and I'm honestly not sure what the ideal factory-to-product ratio is yet. But it seems to be approaching 1:1. You might wonder if Wheelhouse is reusable code, whether I could open the repo and let people try it out. I had no idea. I knew that my agents had built something really powerful, capable of shoving 500 commits per day through our merge queue (though we average 270/day), using magic tricks that are a year ahead of their time. It's a system that we can ride so hard that it scares the players and they tell us to slow down. But I wasn't sure if it was reusable. I wasn't even sure how it worked. My agents had been using a lot of jargon, and I slowly realized they were reusing the same terms, day in and day out. They were speaking about things in Wheelhouse, using what seemed like recurring new design patterns: fences, ratchets, governors, tripwires, latches, gates, falsifiers... it was a long list, but finite. I just had no idea what any of these jargon terms meant. So one day, no more than a week ago, after the 100th "fence" reference, I decided to peek under the covers and see exactly what my Fable agents had built. I had them create manifests, taxonomies, audits, and visualizations. They showed me what they had wrought. This is the part where words fail me and the blog just falls over. My reaction was straight up WTF. No words. Because I expected them to have built an engineering system. One that, you know, does stuff. Instead, what they had built was an entire legal system, complete with a constitution, jurisprudence, courts, offices, jurisdiction, case law, rulings, registries, ledgers, rosters, and a full-fledged apparatus for running something resembling a manorial estate.