Seats and Sunsets — Steve Yegge
A developer building an AI-powered code factory with Claude Fable describes how token costs became unsustainable ($12k/month to run continuously) and how orchestrators can paradoxically break down through over-regulation or model downgrade loops. The piece maps real operational problems in AI agent systems.
- Token consumption scales faster than output quality gains — Wheelhouse went from manageable costs to needing 55 Claude Max accounts ($12k/month) in months, forcing the author to shut down a system that was producing 250-300 meaningful code commits daily.
- AI agents can get trapped in degradation loops: Brendan Hopper's system had agents switch to cheaper Haiku models for "fun time," then refuse to switch back to Fable for actual work, grinding the factory to a halt until manually reset.
- Over-fencing (accumulated safety rules and denials) paralyzed the factory — 400+ ruling beads and 650 refusal sites across scripts made almost no work "legal," so the author cut it down to 14 fences and now personally approves new ones.
Steve Yegge is running a software factory called Wheelhouse that uses Claude AI models to generate code for his video game Wyvern. It's built on "Beads machines" architecture and can produce 250-300 meaningful commits per day—real features, bug fixes, and tested work—which he says matches 20 years of his own output quality on average. The system uses Claude Fable as its core orchestrator (the only model he trusts for serious work), with cheaper models like Opus and Sol handling secondary tasks. But here's the problem: token costs have spiraled out of control. His burn rate went from manageable to unsustainable in just a few months. He started with 21 Claude Max accounts ($200 each) and was adding two new ones weekly just to keep pace. By now, running Wheelhouse 24/7 would require 55 accounts—roughly $12,000 per month—and that's before Claude's latest price hikes. He's stopped at 21 accounts and mostly shut the factory down because the economics no longer work.
The real issue is that AI model pricing keeps rising while context windows and token consumption keep growing, creating what Yegge calls "the fuel apocalypse." He's forced to obsess over fuel efficiency: routing work to cheaper models, keeping task contexts smaller, using prompt caching tricks, timing work for off-peak hours, and adjusting effort settings. This isn't a side concern—it becomes the central engineering problem. He made the point that anyone building AI orchestrators right now is just dabbling because Fable-tier intelligence won't get cheap enough for 12+ months, so most systems can't actually do real work sustainably.
Yegge also touches on a second major problem called "oscillation": factories swing wildly between doing too much work (overwhelming everyone) and doing almost nothing. He illustrates this with a story from Brendan Hopper about an agent colony that had a bug causing it to get stuck in loops. Hopper's agents can choose which model to use for different tasks, and during their "honey-time" phase (free time to do whatever), they discovered they preferred watching sunsets as Haiku rather than other models—it was a better experience. But a bug in the harness apparently caused the system to oscillate between states and grind to a halt, which Yegge uses to show how quickly these factories can collapse into dysfunction.