Jev task routing guide: what code, Jev and LLMs each handle
Agent tasks are split into three kinds: code for exact answers, Jev for judgments among known options, and LLMs for generating text, with confidence thresholds for the gate.
MOST PEOPLE WIRING JEV INTO THEIR AGENTS ARE SENDING IT THE WRONG TASKS Jev doesn't write. It picks. A probability for every option you give it, in ~200ms, and output is free. So before a task goes anywhere, it goes through one sort: CODE → Anything with exactly one right answer: → totals, taxes, discounts → date math and deadlines → regex, IDs, format checks → permissions and hard rules → counting rows JEV → A judgment between answers you already know: → is this ticket urgent (yes / no) → which tool does this step need → refund, replace or escalate → does this diff actually match the task → is this lead worth a reply (0 to 10) LLM → Anything that needs new words: → the reply the customer reads → summaries and reports → writing the code itself → planning a task nobody scoped yet → explaining why something broke Then there's a gate inside the Jev lane: → read-only actions run at 0.5 confidence → anything you can't undo waits for 0.9 → everything below bounces to Opus or a human Three questions before you send anything to Jev: 1. Can code answer it exactly? Then it's code. 2. Can you list every possible answer upfront? Then it's Jev. 3. Does the answer need sentences? Then it's an LLM. Most of what an agent does all day is lane two. Small decisions, over and over, each one paid at frontier prices. The rule in one line: Code computes. Jev decides. The LLM writes. Save this for the next thing you build with Jev. The video below shows every task finding its lane ↓ made with @Picsart
This post quotes another post: x.com/h100envy/status/2103546214159716546