| name | consensus |
| description | Use when a question has a factual or best-practice answer but a single response might hallucinate, or when the user asks to "get consensus" or "sanity check this with multiple agents". Spawns N subagents on the same question with varied personas and tallies answer frequency. |
Stochastic Consensus
One model answer is a sample. Five are a distribution. Frequency across independent samples filters hallucination: a made-up API appears once, a real one appears five times.
Procedure
- Fix the question. Compress it to one self-contained prompt with all needed context inlined. Subagents share nothing; each gets the full question.
- Spawn 5 subagents in parallel (3 for cheap questions, 7 when the stakes justify it). Same question, varied personas, one line each: "answer as a skeptical staff engineer", "answer as a security reviewer", "answer as someone optimizing for simplicity", "answer as a maintainer who will own this for 5 years", "answer with no persona". Personas decorrelate the samples; the question never changes.
- Require structured answers. Each subagent ends with
ANSWER: <one line> plus brief reasoning. Unparseable answers get discarded, not interpreted.
- Tally. Group equivalent answers, count frequency.
Report
- Consensus: the majority answer and its count (e.g. 4/5).
- Outliers: every minority answer, verbatim, with one line on why it diverged. Outliers are not noise. A 1/5 answer is usually a hallucination but occasionally the only agent that spotted the real problem. Flag any outlier whose reasoning cites something specific the majority ignored.
- Confidence: unanimous = act on it; 3/5 split = the question was underspecified, say what's missing; no majority = report that, do not fake a winner.
Cost honesty: 5 subagents is roughly 5x the tokens of asking once. Use for decisions where being wrong costs more than the tokens.
