Not the sales page
The 10-step setup guide
How we sort work before an install. If you want this done on a live fleet, start with an Agent Leak Audit.
01 — Find what Jev can take over
Creates text? LLM. Picks, scores, or gates? Jev. Fixed numeric rule (stop after 10 loops)? Deterministic code.
02 — Design the state Jev reads
Hand it goal, evidence, completed work, and constraints. Vague state produces vague decisions. Jev does not go looking for context.
03 — Test one question before wiring code
Ask a single Choice — “Which worker should act next?” — on real state. If the answer is wrong, sharpen the state, not the poetry.
04 — Learn the three question types
Choice routes (1 of N). Score ranks. Noul gates yes/no. Every leak maps to one of these.
05 — Funnel a large list to one
Code eligibility → Jev Score survivors → Jev Choice winner. Do not dump 200 tools into a frontier prompt.
06 — Dispatcher / chief-of-staff
Jev sits between shared state and workers. Confidence ≥ your threshold: dispatch. Else: human_review.
07 — Batch questions on the same state
Worker choice, urgency, and approval can be one round trip. TypeSafe documents parallel questions on one request.
08 — Harness: router on top, gate on the bottom
Model router picks the cheapest competent model. Auto-mode gate checks destructive tools before execution.
09 — Compact with Score, do not summarize
A summarizer rewrites history. A relevance filter keeps or drops. Keep generation for generation.
10 — Count the cost you removed
At TypeSafe’s published $0.042/M input and $0 output, 10,000 × 1,000-token decisions are about $0.42. The homepage ROI teaser is the napkin math; an audit uses your traces.