← all entries
A-TIER technique

Agentic OS on Claude Fable 5 — cron loop with earned autonomy

raycfu · 2026-07-10T02:02:12.000Z · in AI Saves, Vol. 8
Open post ↗

100k ft — What is this?

Agentic OS on Claude Fable 5 — cron loop with earned autonomy
techniqueA-tier

50k ft — Why care?

Blueprint: a cron-run loop.sh where a cheap model triages changes, Fable 5 plans as manager, Sonnet executes and a fresh Fable reviews. Tasks earn autonomy after 20 passes at 95% and lose it below 90%, refusals come back as HTTP 200 so scripts must check stop_reason, and a goals directory reruns regression tests daily.
LLMagentssoftware developmentvideo production

10k ft — Key takeaways

• Check stop_reason for refusals, not just exit codes • Grant autonomy per task type from logged pass rates • Keep daily regression tests so old builds can't break silently

5k ft — Tools & techniques

Claude Code, cron AI agents, earned autonomy, manager-worker-reviewer, model routing

1k ft — Deep-dive / try it

How to build an Agentic OS using Fable 5 part 1 #ai #tech
Transcript excerpt
Most people don't realize that Cloud Fable 5 can run for hours unintended, dispatch its own subagents, and do a whole week of work in a single night. Here's how to build an actual agentic OS using Fable 5. Start by configuring the agentic OS engine. The model string should be set to Fable 5, and make sure to set max tokens to 64k, so it caps the thinking plus the responses. Effort is always set to low to max, but never extra high unless you're working with an agentic loop. And here's the part most people miss. Refusals need to come back as HTTP 200. In this way, your scripts have to check for stop, reason, refusal, and not just exit code. Create a batch script called loop.sh and it runs on a cron drive every morning and runs your entire system. First, use a cheap AI model like Gwent to look at your project. And ask if anything changed or needs attention. If yes, Fable steps up as the manager. It does not do the actual work, but writes down exactly what needs to be changed. Then use Sonnet to do the actual work, and a fresh copy of Fable that checks it like an inspector. Next, build up the report card. Every task the system finishes gets logged as a pass or fail, and each type of task will earn its own grade over time. Once something passes 20 times with a 95% success rate, the system is then allowed to do it completely on its own. But if the grade ever drops below 90%, the system will lose that privilege, and you get an email. So this way you never have to guess whether to…

Appears in

More from @raycfu

Collection: AI Saves, Vol. 8