Four AI workers fix bugs together: a researcher, a coder, a tester and a reviewer. Something has to decide who works next, and in most frameworks that is another large model re-reading the whole conversation on every turn. Here Jev makes that choice in a quarter of a second. It also checks whether the last piece of work was any good, and notices when the team is going round in circles.
Press Play. Each step is one Jev call returning four answers at once.
Raise the confidence floor to 0.7 and the first bug goes to a person even though it was going fine. Lower the step budget and the flaky test is stopped by the budget before Jev spots the loop. Changes apply from the next decision.
| Task | Path | Outcome |
|---|
The numbers are illustrative, not measured: 0.25s and about $0.0001 per Jev call; an LLM supervisor at about 4s and $0.01 per call, more as the transcript grows. ← 12 · Four specialists, one verdict · Next: 14 · Passing the customer on →