Explore a problem from multiple angles simultaneously. Spawns N sub-agents with different framings, collects answers, and ranks by agreement. Use when the user says "ge mig approaches", "utforska alternativ", "consensus", or "vad finns det för sätt att...".
Stochastic consensus — like asking multiple experts instead of 1.
Understand — User describes a problem or question. Determine output type:
Generate framings — Create N variants (default 5) with different perspectives:
Define structured output — Before spawning, define what each agent must return. Must be mechanically aggregatable:
Spawn sub-agents — Run all N in parallel via Agent tool with fresh context. Each gets: framing + problem + context + output schema.
Aggregate — Method depends on output type:
Ranking: Assign points (1st = N, 2nd = N-1...), sum across agents, report final ranking.
Recommendation: Group similar ideas (fuzzy match), count how many agents proposed each:
| Agreement | Category | Action |
|---|---|---|
| 80%+ of N | CONSENSUS | High confidence. Recommend directly. |
| 40-79% of N | DIVERGENCE | Genuine tradeoff. Present alternatives, let user decide. |
| <40% of N | OUTLIER | Potentially brilliant OR hallucinated. Flag for review. |
Scoring: Calculate mean, median, and standard deviation per option. Flag options with std dev > 2 (agents disagree).
Binary: Count YES vs NO. Report the split and summarize strongest arguments from each side.
Present — Summary with: