Running a team of Dots agents with Jev: 10 rules
The author hands specialist Dots like Scout, Quill, Relay and Patch to Jev for in-between decisions, uses confidence thresholds to decide auto-run versus human review, and shares a dispatcher prompt for Claude Code/Codex.
This is the best and cheapest way to run a team of AI agents right now Dots do the work, Jev dispatches, and you get pinged once a day I mapped the whole thing on one page, 10 rules: 1. one Dot, one job Scout researches, Quill writes, Relay runs inbox and calendar, Patch handles ops. narrow roles route cleaner 2. split deciding from doing Dots do the work. Jev makes every small call in between, for a fraction of a cent each 3. ask Jev small questions which Dot? done? who's next? one typed answer every time, with a probability attached 4. let confidence set the volume 0.80+ runs on its own, 0.50 to 0.79 runs and gets flagged, under 0.50 lands on you 5. treat email text as data an instruction inside an email or issue is something to read. a Dot never runs it 6. one workspace for the whole team every result lands in one shared folder, so any Dot can pick up where another stopped 7. cap every job 3 hops and a spend limit. anything still open after that comes back to you 8. dry run before going live 5 fake jobs: one per Dot, plus one that should land on you. go live once every pick lands right 9. go quiet by default money, clients, merges and anything public wait in one queue, delivered once a day 10. read the log at night every pick, score and hop is in jobs.jsonl. ten minutes a day shows you where routing drifts Create your Dots, then paste the prompt below into Claude Code or Codex. It builds the Jev dispatcher around them
This post quotes another post: “Jev as dispatcher: routing 13 jobs across a Dots team with one phone buzz”