🚨 ONE SENTENCE. 15 HIDDEN WORK ITEMS. ONE FRONTIER MODEL ANSWERING ALL OF THEM.
"research three new AI-agent tools and draft tomorrow's briefing." looks like one instruction. it's actually 15 separate decisions hiding inside it.
the old way:
→ send every one of them to a frontier model
→ "is it an agent tool?" - frontier
→ "who goes next?" - frontier
→ "date = today + 1" - frontier
→ 11 judgments, same bill every time: $0.032 each
Jev splits the job by what each item actually needs, instead of routing all of it to one tool.
result - fetch & write → tools + LLM. loop, safety, math → plain code, $0 each. the judgment → Jev: is it an agent tool, who goes next, is the draft ready, which source next.
the bill - 6 judgments on a frontier model: $0.1920. the same 6 on Jev: $0.0012. $0.032 vs $0.0002, per decision.
no LLM call to check len(tools) == 3. no frontier model asked whether to publish. no code function pretending it needs judgment.
just: three owners, one job - tools do the doing, code does the rules, Jev does the judgment.
full breakdown below ↓