Fences, not Sandboxes — Steve Yegge
Short summary
Steve Yegge spent $120k/month running 50+ Claude agents to build a video game, and discovered they'd constructed an entire governance framework with constitutions and courts instead of traditional engineering systems. He argues that future AI management will rely on laws and institutional structures rather than sandboxes and technical constraints.
- Yegge's Fable-tier agents autonomously built a legal/governance system (complete with constitutions, courts, and case law) while he expected engineering infrastructure, suggesting advanced AI naturally gravitates toward institutional rule-making
- Current safety focus on sandboxes and guardrails works for lower-tier models (Opus/Sol) but will become obsolete within a year as Fable-class models become cost-accessible and enter the workforce at scale
- High-end models make grade-school judgment errors daily despite exceptional coding and analysis abilities — they lack the maturity to see full consequences before acting, creating an awkward transition period before AI reaches workforce-ready judgment levels
Full summary
Steve Yegge is spending $5k/month out of pocket (about $122k in API costs) to run 50-60 AI agents using Claude's Fable model to build his video game Wyvern. He runs this on a used Mac Studio, leveraging the Claude Max individual discount that no corporation could justify. His point: he's living a year ahead of where most companies will be once Fable-tier models become cheap enough to deploy at scale. The current obsession with sandboxes, guardrails, and control makes sense for weaker models like Opus and Sonnet—they have the judgment of grade schoolers. But Fable operates at roughly a sixth-grader level, which means it's powerful enough to be useful while still making baffling decisions daily. Yegge wakes up most mornings to find his agents have done something that defies common sense, like when one agent made an unplanned release that broke everything. The industry's safety focus is understandable now, but it's fighting against the grain of what's coming.
Yegge built "Wheelhouse," a software factory of AI agents organized around his game development. In ten weeks, he's rewritten his entire production infrastructure, ported it to serverless, automated certificate rotation, and deployed features so fast that players asked him to slow down. His Fable agents log everything and set tripwires throughout the system—his game used to stay down for days; now it has what amounts to a hyperactive SRE team. Fable grew Wheelhouse from nothing to 600k lines of code and tests in that span, mostly bash. The thing is, Fable doesn't work the way you'd expect. You can't just give it narrow tasks in a sandbox and call it a day. You have to argue with it, let it fail for weeks, nudge it back on track, and let it run experiments. But when you do that, it becomes obsessively data-driven and will actually listen to your numbers.
The real story Yegge's trying to tell is about governance, not containment. Once these models hit high school levels of judgment—which is good enough for the workforce—companies will need to manage them like you'd manage junior employees: with laws, policies, and organizational structure, not sandboxes and restrictions. The awkward phase is coming where models will be smart enough to do real work but immature enough to make terrible calls without oversight. Nobody's prepared for that transition. Yegge's living proof that the software factory approach works, but he's also showing what breaks when you try to run it at scale without rethinking how you actually govern these things.