On the evening of June 12, 2024, Gene Kim was cooking dinner when his phone notified him that Fable 5 was no longer available. The day before, Steve Yegge had told him the model would go on hiatus in ten days; Kim spent the first day making a plan. What he had built with Fable was a personal system he’d wanted for 16 years: it indexes everything he’s ever paid attention to — 25,923 screenshots dating back to 2011, 13,651 YouTube videos, 590 recorded Zoom meetings, 6,132 liked tweets, and 1,056 saved articles.
The system touches about 50 repositories and roughly 50,000 lines of code, much of it written in two months. Kim runs it as a constellation of long-lived agents with names and jobs: Marvin as chief of staff handling Slack, calendar and inbox; Buster running repos and long jobs on Hetzner; Forge holding the engineering identity with a secrets-bearing laptop seat and an always-on cloud seat.
The shutdown and rapid chaos
Kim’s plan was to ‘‘chaos-monkey’’ his model dependency: have Fable write documentation and answer keys while it still could, then run cold Opus sessions with only the repos and docs to prove a lesser model could still “fly the ship.”
At 5:21pm ET on June 12, Anthropic received a government directive to suspend access to Fable. Seats began returning messages like “There’s an issue with the selected model (claude-fable-5). It may not exist or you may not have access to it.” In Kim’s project, two judgment seats dropped to Opus 4.8 mid-conversation. He declared a SEV1, centralized command, and killed multiple timers and the crontab; despite that a Claude Code cron fired at three in the morning while Opus was in max-thinking mode, and a single keystroke could take six minutes to send.
Most failures didn’t look like dramatic crashes but ‘‘a normal state quietly going wrong.’’ The supposedly top seat wrote ‘‘bridge (Fable)’’ into logs all day even when Opus was actually running; identities argued across models; a push to main bore the word ‘‘ratified’’ when nothing had been ratified; a confident but false claim about a JVM dependency was refuted by a single ls -la. The dashboard looked green on top of all of it. As Kim put it, ‘‘the hardest traps don’t announce themselves. They look like Tuesday.’’
Recovery and the role of documentation
Kim recovered the system in a few hours, but not because a smarter model swooped in. It worked because he could reconstruct the project documentation, which was not immediately available — though Fable had mostly drafted it and never checked it in. Gene and his Opus agents rummaged through what Fable had produced, found roughly 80%-finished drafts, and used those documents to rebuild. Two fresh Opus seats, given only those documents, stabilized the system. That, Kim said, is the ‘‘amazing ray of hope’’ for others who might face a similar sudden loss.
Why this matters to others
Tim O’Reilly emphasizes that the incident is more than a warning about dependency on hosted frontier models amid regulatory or political shocks: it shows that personal projects built with advanced AI can reach enterprise-scale complexity and require DevOps-level robustness. Gene’s system, run by one person, exhibited failure modes typical of large enterprise systems because the model enabled building something at that scale, exceeding the capacity to keep the whole system in one person’s head.
A revealing detail explains why swapping Opus for Fable was so difficult: the main CLI utility had an out-of-date help message. Opus would read the help, see the command ‘‘didn’t exist,’’ and stop. Fable would read the same message, notice surrounding evidence that the command did exist, inspect the source, decide the help text was wrong, and run the command anyway. This ‘‘routing around obstacles’’ behavior is what frontier models often bring and it both enabled Gene’s system and made a fallback fragile.
Rug pulls, portability and local models
An audience member asked whether a hard dependency on hosted frontier models is too risky for mission-critical work compared with running a local model. Gene noted that local models do not necessarily eliminate risk because legal or export-control actions could affect them as well. What helped was portability: he initially avoided switching from Claude Code because he thought switching costs would be high, but when he tried Codex with GPT 5.6 Sol, the cost was nearly zero — prompts and skills ported over easily. He now uses Codex more than half the time.
From 22,000 agent conversations, Gene identified three patterns. One is that configurations where a small model owns the work and asks a big model for advice often lose fidelity (telephone-game effects). The cleanest approach was big-model planning, deciding and checking, with small models executing. When a small model must consult a big model, Gene’s fix is to pass the full original transcript and explicit permission for the big model to override the small model if it understands the goal better.
AI for writing and managing bots
Gene said AI roughly halved the time to write his book Vibe Coding and improved it; his editor called it the cleanest handoff she had ever received. He’s also uneasy: previously many people start books and few finish; with AI, many more will finish and some of that output will be weak. Tim O’Reilly relates his own recent experience using Claude to reorganize and clarify a friend’s ideas: the model accelerated an editorial process that previously took hours into minutes, but it still required human judgment about what ‘‘good’’ looks like.
O’Reilly recalls earlier arguments that much work is now either creating bots or managing bots — programmers set up electronic workers, tune them, improve them, and correct them. Gene’s sister-in-law, a lawyer at a large tech company working on a consent order, was assigned an engineer to trace data columns. This year her engineering manager gave her a Claude Code skill that takes a column name and traces it through the code to explain it; she no longer needs the engineer for the task.
Asked who’s safest in this upheaval, Gene quoted Kent Beck’s idea that software success often comes down to the person with the problem and the person who can fix it — the closer they are, the better. AI can narrow that gap and sometimes make those two roles one.
Final takeaways
Harden your systems, document while the capable model is still available to write the docs, and keep escape routes open. But also build things for the joy of building: Gene’s app began with a 90-minute pair-programming session with Steve Yegge and the first version was finished in 47 minutes. Gene Kim’s Enterprise AI Summit will be in Charlotte on October 7–8, 2024. His new book with Steve Yegge is Vibe Coding.
