001 - Rysh Use Cases - Loop Engineering with Rysh

Watch an AI agent approach 80% of its budget and hand the wheel to a wrap-up prompt — instead of dying mid-task or burning past its ceiling. That's the 80 seconds in this video, and nothing in it is a mockup: a real terminal, a real automation, a real budget. The #1 blocker to letting agents run unattended isn't capability. It's control. Here's what control looks like in rysh: — A hard token budget in units a human can say out loud: this run gets twenty pages. (A page is 1,000 tokens; 200 pages make a book.) Enforced, not suggested. — A takeover step at 80% spent: a wrap-up prompt inherits the reserved 20%, saves results, labels them PARTIAL if unfinished. Nothing dies mid-thought. — An LLM judge with a goal you wrote down decides "done" or "go again." The stop condition lives in the diff, not in someone's head. — Stop any time; continue resumes from the last checkpoint. — The whole loop is a YAML file in your repo. Reviewable, versionable, owned. Then you schedule it with cron and go to sleep. That's the difference between "we tried agents" and running them like any other production workload: observable, governed, owned. We're early, and onboarding design partners now. Comment "leash" or DM me and I'll show you a live run in 15 minutes.