Jev as decision layer: breaking down a 24/7 agent bill
The author uses Jev to answer three questions (lane, urgency, evidence) per event, finding that 97% of costs come from the 4% of events needing writing while Jev accounts for under 3%.
Jev tip nobody tells you: once decisions are nearly free, your agent bill lives in your escalation rate "a frontier LLM charges $600 a day for 20,000 decisions, Jev does them for $0.84, and the whole desk still costs $29.04" every post, email, alert and issue goes through Jev before anything else touches it Jev picks, scores and says yes/no for $0.042 per 1M tokens, and output is free it can't write a single word, and it can't answer outside the options you give it that limit is why it works as a decision layer for a 24/7 agent Grok Bot only gets paid when an event actually needs something written each event gets 3 questions in one request, answered in parallel: → lane: ignore, log, draft or human? → urgency: later, today or now? → evidence: are the facts in the event, or missing? the 3 questions can't see each other, so you get a free second opinion on every event when urgency says "now" and lane says "ignore", Jev is confused, and the event goes to your queue for zero extra calls - where the money actually goes → 97% of the bill is the 4% of events where Grok Bot has to write → Jev is under 3% of the total, so most people are optimizing the wrong 3% of their bill → the 1% audit of ignored events costs $4.20 a day and is the most useful line on the bill now do what everyone does once decisions feel free: watch 10x more sources and ask 5 questions instead of 3 the bill jumps from $24.84 to $251.76 a day, with $10.92 from Jev and $216 from writing it breaks time too, because at 500ms worst case 200,000 decisions take 27.8 hours on one worker your "24/7 desk" can't finish one day in one day, so plan for two workers before you scale - the full setup → Jev pinned to jev-1.13.0 on every event, never a moving alias under your threshold → event text labelled as content in every request, so instructions inside it never run as commands → every Jev error, timeout or missing key goes to human, never quietly into ignore → lane confidence under 0.85 goes to human → urgency "now" with lane ignore or log means the questions disagree, so human → draft with missing evidence goes to human, because a draft without facts is a guess → 1% of ignored events get audited forever, because nobody else ever looks at that lane again → every event leaves a line in ledger.jsonl with its answers, lane and token count - what raising the threshold actually does → your first instinct is to raise the confidence threshold "to be safe" → the jev-harness repo swept it from 0.50 to 0.90 on a coding-agent gate → zero bad proposals got through at any level, because Jev's answer did the catching → the threshold on its own only blocked good proposals and piled them on a person → so treat FLOOR as a cost dial, and check the audit lane before you lower it → change one thing a week and read the audit after every change - Grok Bot runs two jobs on one cloud box → Desk Bot keeps the loop alive and drafts every 10 minutes into a review queue → Ledger Bot sends the bill at 09:00 and puts a warning first if escalation passes 6% → Grok Bot never picks lanes, never approves its own drafts and never posts anything $42 buys a million decisions on Jev, and after that your bill follows the escalation rate save this before you build your next agent, then hand the tree and this prompt to Claude Code ↓ "Read ledger.jsonl from my jev-desk and tune it against this tree: 1. Print the daily bill from the ledger → Jev cost at $0.042 per 1M tokens → writing cost at $0.03 per draft and per audit → Jev share of the bill and the escalation rate → lane split: ignore, log, draft, human 2. Find the questions that disagree → count events where urgency is "now" and lane is ignore or log → group them by source and show 3 examples from each 3. Replay the logged lane confidences against FLOOR from 0.70 to 0.95 in 0.05 steps → show how many events each value sends to human → show how the daily bill and my review load change at each step 4. Check the silent lane → list every audited ignore marked as a false ignore, grouped by source → estimate the false ignore rate from the 1% sample 5. Project the bill at 10x events with 5 questions per event, and tell me how many workers I need at 500ms per call 6. Tell me which single change cuts escalation the most without raising false ignores Check the ledger field names against the real file. Show me the numbers first. Don't change any code yet."
This post quotes another post: x.com/gippp69/status/2103394701063737504