When Fable5 Dropped, I Spent $1000 on Tests
I had it build full games from a single prompt. I had it create visualizers from a single idea. Meta-orchestration seems to be working very well…
Fable is the sanitized, consumer-safe cut of something Anthropic calls “Claude Mythos” which was recently walled off to the US government and told that plebs like me can no longer have access. Do I get a refund?
I run Spec-Driven Development: a human writes a tight technical spec with ai, then you turn the agents loose. This time I broke my own rule. One vague prompt. No spec. A job that needed research across thousands of permutations of actions, quests, real fighting and action decisions, and hundreds judgment calls along the way to create a narrative and storyline....
In under an hour, Fable researched it, planned it, built it, and tested it. No obvious bugs. It wasn’t pretty. But it worked… and it was better than what my 4 agents created previously…
It Was Running a Team
The output was kind of expected. T
Mid-build, Fable stopped and asked me four questions. Not clarifying-the-prompt questions, but complex decision questions. How many talents on the skill tree and leveling up cap? NPCs and their roles within the game. Side quests and main quest requirements, and progression mechanisms. These questions were the Product Manager reading between the lines of a thin brief and knowing the domain enough to ask the questions…!
So I went and pulled the build session logs. Fable was running a team of 12 sub-agents running on less capable models like Sonnet or Haiku, and it was managing all of them itself:
4 Business Analyst agents sent out across the internet to research game mechanics and gamification.
4 Builder agents to create the scaffolding, architecture, logic, design…
2 Reviewer agents spun up specifically to review the architecture and the code PRs… basically acting as tech leads.
2 QA Tester agents dropped into a real browser to test the finished app with one running a headless to do screenshot testing from different angles…
Coordination, orchestration, hundreds of small calls… all of it off a single high-level prompt from me.
Where It Falls Down
It’s not magic, and it’s not free.
It’s fabulously expensive! Twice the cost of Opus. I burned through hundreds of dollars of tokens on this one adventure…
The guardrails are hair-trigger. Fable would hard-shut a perfectly legitimate conversation the second a prompt brushed against cybersecurity, bioengineering, even applied physics. Not “let me reframe that".” Full stop. These were easy to test.
Conductor to (Paying) Patron
I’ve been watching the unit-economics on tokens very closely. I have a dashboard of token use from the 8 machines I’m running + a bitcoin miner :)…
AI started as a power tool. A better tool. You still had to use it.
Then we became conductors. Orchestration of agents. Still reading the score… still watching the progress.
With Fable, I’m now a paying patron.
I commission the work. I sit down. Watch the show. The orchestra makes thousands of decisions I never see and cannot fully audit (have you looked at this llm code? im’possibruuu!!).
Strip it all the way down and yah… it might still be a very sophisticated parrot guessing the next token. But functionally, it is performing at superhuman levels in software engineering. An immense cognitive tool, unbound by conscience or anything resembling humanity.
And the honest, uncomfortable part: I’m starting to lose the ability to even comprehend what it’s doing… let alone control it.
Best,
ps