society-of-mind

$npx mdskill add yogsoth-ai/de-anthropocentric-research-engine/society-of-mind

Facilitates multi-agent collaborative debate to refine perspectives iteratively

  • Solves complex problems by simulating diverse viewpoints and resolving disagreements
  • Uses subagents like debate-architect, perspective-critic, and divergence-detection
  • Decides based on iterative sharing and revision until convergence or divergence
  • Delivers a synthesized verdict through debate-transcript-analysis
SKILL.md
.github/skills/society-of-mindView on GitHub ↗
---
name: society-of-mind
description: "Strategy: Multi-agent collaborative debate based on Du et al. Society of Mind. Agents share perspectives iteratively until convergence or divergence is detected."
type: strategy
used-by: [multiagent-debate]
tactics: [perspective-rotation]
---

# Society of Mind Strategy

Multiple agents evaluate independently, then share and revise through iterative rounds.

## Method

1. **debate-architect** assigns N distinct perspectives to agents
2. Each **perspective-critic** evaluates artifact from assigned viewpoint
3. All perspectives shared — agents revise positions given others' arguments
4. **divergence-detection** identifies agreement clusters and persistent disagreements
5. Repeat sharing rounds until convergence or saturation

## Budget Table

| Parameter | S | M | L |
|---|---|---|---|
| Debate rounds | 4 | 8 | 12 |
| Participating agents | 3 | 5 | 8 |
| Coverage dimensions | 3 | 5 | 7 |
| External evidence searches | 2 | 5 | 10 |

## Orchestration

```
debate-architect → [assign perspectives]
→ [parallel]: perspective-critic × N
→ [for each sharing round]:
    share all outputs → perspective-critic revises
    → divergence-detection → (converge or continue)
→ debate-transcript-analysis → verdict-synthesis
```

## Subagents

- debate-architect (perspective assignment)
- perspective-critic × N (perspective evaluation)
- divergence-detection (convergence tracking)
- confidence-calibration (termination decision)
More from yogsoth-ai/de-anthropocentric-research-engine