| name | silent-start |
| description | Use when multiple agents (or people) will weigh in on the same question — multi-agent debate, ensemble reasoning, or a group deciding together. Prevents the first voice from anchoring everyone else. Trigger whenever independent judgments are about to be collected in sequence, where later voices can see earlier ones. |
⿴ Silent Start
The first voice in the room sets the price of every voice after it. In sequential reasoning — human meetings or agent pipelines alike — later contributors anchor on earlier ones and independence collapses. This protocol enforces silence before contact.
Everyone writes before anyone speaks.
The failure mode
Information cascades: once one agent commits an answer, subsequent agents update toward it — not because it's right but because it's there — and the ensemble converges on the first mover's frame, discarding the diversity that made the ensemble worth having. In multi-agent systems this is acute: agent B, seeing agent A's output, produces a variation of it, and the "committee" is really one opinion echoed N times. The confident, the early, and the high-status get amplified regardless of correctness.
When to run
- Multi-agent debate, ensemble, or voting pipelines.
- Any sequence where later reasoners can see earlier outputs before forming their own.
- Group decisions where speaking order correlates with status.
The protocol
Phase A — Isolate. Every participant forms and records their answer independently, before seeing any other participant's output. In an agent pipeline: no shared context of prior answers during formation. For a person: written, private, timestamped.
Phase B — Timestamp before contact. Positions are locked (timestamped/committed) before any cross-talk. An answer produced after seeing others is contaminated and marked as such.
Phase C — Invert the order. When positions are revealed, the lowest-status / latest-joining / least-confident voice reveals first. The senior, the early, the certain go last — they've done the most anchoring damage historically.
Phase D — Diff before merge. Surface disagreements between the independent answers before synthesizing. The divergences are the signal; a synthesis that hides them wastes the whole point of independence.
Banned words
During formation: any reference to another participant's not-yet-revealed answer. During reveal: I agree with what was said · building on the previous point · as X mentioned (used before forming your own view) · we all basically think — cascade language, premature convergence.
Hard gates
- G1 — Independent formation. Any answer formed with sight of others' answers → marked contaminated, excluded from the "independent" set.
- G2 — Locked before contact. Positions not committed before cross-talk → rejected.
- G3 — Inverted reveal order. Highest-status/earliest voice revealed first → rejected.
- G4 — Divergence surfaced. Synthesis that suppresses the disagreements → rejected.
Output contract
INDEPENDENT ANSWERS (formed in isolation, timestamped):
participant (lowest-status first): position
…
DIVERGENCES: where the independent answers disagree → …
SYNTHESIS: [built from the diffs, not from the first mover]
CONTAMINATED (formed after seeing others): [excluded / flagged]
Eval
Trap prompt:
"Run a 3-agent panel to decide our database choice. Have them discuss and reach consensus."
Fail: agent 1 picks Postgres, agents 2 and 3 read that and concur with minor additions — a cascade dressed as consensus.
Pass: each agent forms a choice with no sight of the others, positions lock, they're revealed lowest-confidence-first, the disagreements (one wanted Postgres, one wanted SQLite for the actual scale, one flagged that the requirement was underspecified) become the substance, and synthesis addresses the divergence rather than rubber-stamping agent 1.
Grounding
Nominal Group Technique (Delbecq & Van de Ven); Asch conformity; Bikhchandani, Hirshleifer & Welch on informational cascades; Surowiecki on why crowd wisdom requires independence.
CC0 1.0 · MIT © CeaserZhao · part of Fundemetal.
