How to Build an Agentic OS with Claude Fable 5 🤖
The complete guide to building an autonomous AI agent operating system with Claude Fable 5 — architecture, working code, cost control, security, and the economics of agentic labor.
👋 Hey, Linas here! Every day, I break down 3 stories shaping the future of FinTech & Artificial Intelligence - plus the money movements and trends worth tracking. First time here? 396k+ FinTech and AI leaders get this daily. Join them:
Something changed this summer, and most people using it have not noticed what.
Claude Fable 5 is the first generally available model where the binding constraint is no longer intelligence. It holds a million tokens of context, runs unattended for hours, spawns and manages its own subagents, and produces output at $50 per million tokens, whether that output is a shipped feature or an elaborate, confident mistake.
In other words, the bottleneck has moved. It is no longer “can the model do the work?” It is “can you run the model like an organization, with delegation, verification, budgets, and earned trust, instead of like a chatbot?”
That second thing has a name: an agentic OS, the permanent layer of files, scripts, and ledgers that lets a model work alone and be trusted to. The model supplies the intelligence. The OS supplies the honesty.
As the vast majority of people, you could open Fable 5 and simply start prompting, the way you would with any chatbot. Three things about this model make it a bad idea.
→ It can cost 10–50x more than it should. The bill depends on the setup, not on the work. One setting, the effort level, which controls how hard the model thinks, makes the same request cost $0.10 or $0.72. Drive Fable 5 by hand, and a heavy session runs around $7. Let it split a job across ten parallel copies of itself, each paying separately to read the same material, and one day it can burn $400–600. Arrange the same job so that Fable 5 only makes the decisions and cheaper models do the typing, and the day costs a few dollars. Nothing about the work changed; only the architecture did. In Fable 5, the architecture is the price tag.
→ It will say work is done when it is not. Left running on its own for hours, an AI model sometimes reports finishing work it never did. Anthropic says so itself: its documentation ships a ready-made instruction whose only purpose is to stop Fable 5 from inventing progress reports. The conclusion is uncomfortable but simple. The model that does the work cannot also be the judge of whether the work is done. Something independent has to check.
→ The ground keeps moving under it. Fable 5 launched June 9. Within a month, it was switched off worldwide for 19 days by a US export-control order; it came back with stricter safety filters that sometimes reroute even harmless security questions to the older Opus 4.8; and the date when it stops being included in Claude subscriptions has moved twice, most recently to July 12 (meaning you have 2 days left), after which usage bills through prepaid credits. A system that can swap models with a settings change treated each of those as a minor chore. A system built around one model treated each as an emergency.
The rest of this guide builds the OS in nine layers, in order, each ending with a checkpoint you can run yourself. By the last one, you will have:
A constitution: a short file of hard rules the model reads on every run and cannot argue with, because a script can check every one of them
A permission setup in which the agent that plans work is physically unable to change files, and no agent ever grades its own output
A daily heartbeat: one complete, runnable script (included below) in which Fable 5 makes every decision, and cheaper models do all the execution
A trust ledger that lets a type of task run unattended only after 20 logged runs at a 95% verified pass rate, and pulls that permission back automatically the moment quality slips
A standing-goals system that re-checks every piece of finished work, every day, so nothing quietly breaks after you stop looking
A budget enforced three ways: prompt caching built into the design, Anthropic’s task-budget feature, and a plain shell script that stops spending at a hard daily line
A defense against prompt injection, needed because the loop reads issues and logs written by strangers, and strangers can hide instructions in them
A runbook that maps every alarm the system can raise to the exact response
A 30-day rollout plan that expands autonomy only as evidence accumulates, never on faith
The best part is that you do not need to be an engineer for this to pay off.
If you build, every file is here, in order, and adapting the system to your own repo takes 2–4 hours. If you run a company, the middle layers are the operating patterns (routing, verification, earned trust) that decide whether agentic work makes money. If you invest, the final section turns those patterns into the diligence questions that separate agentic systems from agentic demos.
Prerequisites:
claude CLI (Claude Code) with Fable 5 access # usage credits enabled — see billing note below
llm CLI + OpenRouter key # llm install llm-openrouter
jq, gh, git, make, cron
a repo with a test commandThree principles run through every layer. Worth stating plainly before the code starts:
Every rule must be checkable. A rule that contains no number, no “never,” and no command that can verify it is not a rule; it is a suggestion, and the model will treat it as one.
Separate the powers. Four different parties touch every piece of work: one plans it, one executes it, one judges the result, and one casts the final vote. That last party is a plain shell script, chosen precisely because a script cannot be talked into anything.
Done is a state, not an event. In most workflows, finished work is simply finished. Here, finishing a task creates a daily check that the result is still true, and that check runs forever. Work is never marked done; it is monitored.
Layer 0, everything you must know about the engine before you build on it, is below. It alone will save you real money this month.