A Dot operator's three-layer agent stack: Jev routing plus Astra escalation
DOTS holds state, Jev returns decisions with confidence scores in under 50ms, and Astra handles only the 6% of tasks that need real reasoning, with no manual task routing in three months.
A Dot operator has not manually routed a single task in three months. Meanwhile, his agent handles 10x the volume on a fraction of last quarter's budget. He did not optimize the frontier model. He simply stopped feeding it routine work, letting a cheap decision model handle everything that used to drain Astra. The architecture relies on three core components: DOTS: the agent runs 24/7 on OpenAI's new autonomous workers, holding the entire project state between sessions. JEV: every incoming signal goes to a decision model that responds in under 50ms with a confidence score attached. ASTRA: the frontier model wakes up only for the 6% of tasks that actually require real reasoning. Agent frameworks were built when every call had to pass through the smartest brain available, simply because no other model could decide anything. Jev changed that arithmetic. Structured decisions do not require generation, they require classification. And classification is 238x cheaper per token. The hardest part is routing. A bad route can waste more compute than it saves. That is why every path includes a confidence score, and anything below 0.80 escalates automatically. Most tasks clear in milliseconds. The ones that need a frontier model get it. The ones that touch something irreversible wait for human approval before any action takes place. The agent still writes code, drafts emails, and closes tickets. It just stopped waking the $10 brain for every simple decision along the way. The full breakdown, from the Jev router to the Astra escalation rules, is available in the article below.
This post quotes another post: x.com/shmidtqq/status/2108637104016072888