the-jury

Invoke the jury workflow with a decision question and its options. It spawns 3 to 5 juror subagents that vote blindly, deliberate over anonymized positions, and return one committed verdict with a confidence level and next action.

It is domain-agnostic: engineering and architecture tradeoffs, build-vs-buy, vendor choice, hiring, product strategy, data and research design. It is not for simple factual lookups, pure critique without deciding, or writing the solution itself.

Round 1 votes are blind so no juror sees others first, deliberation anonymizes all positions, and every change of mind must cite a specific new argument. If flips look like majority-chasing, the tally falls back to the independent Round 1 aggregate at low confidence.

Yes. Without a parallel subagent runtime, the panel is simulated as sequential role-played passes in one context, but all Round 1 positions must be generated before any position is revealed to any juror, preserving the blindness gate.

The foreman never abstains or coin-flips. Near-ties are decided on the merits: the option with the higher evidence grade that best survived the devil's advocate's strongest challenge wins, with the closeness recorded as dissent and lowered confidence.

Skip it for simple factual lookups, tasks where you only want critique without a decision, or when you need a plan or implementation written. It adds deliberation overhead, so reserve it for contested or high-stakes questions.