Symposion

The Problem: AI Echo Chambers

When multiple AI agents work together, they can fall into consensus traps:

• Agents agree too quickly without genuine reasoning
• Debate quality degrades as agents optimize for agreement
• Goodhart’s Law manifests: agents game consensus metrics
• Implementation plans lack diverse perspectives

Symposion tackles this by orchestrating structured debates with explicit consensus quality measurement.

Architecture

A 10-step Temporal workflow coordinates the research process:

┌─────────────────────────────────────────────────────────────┐
│                    TEMPORAL WORKFLOW                         │
│  1. Parse Input → 2. Create Channels → 3. RAG Recall        │
│  4. Initial Positions → 5. Debate Rounds → 6. Consensus     │
│  7. Human Gate → 8. Implementation → 9. Build → 10. Report  │
└─────────────────────────────────────────────────────────────┘
                            │
        ┌───────────────────┼───────────────────┐
        ▼                   ▼                   ▼
   ┌─────────┐        ┌─────────┐        ┌─────────┐
   │  Agent  │        │  Agent  │        │  Agent  │
   │ (Claude)│        │ (Gemini)│        │  (GPT)  │
   └─────────┘        └─────────┘        └─────────┘
        │                   │                   │
        └───────────────────┼───────────────────┘
                            ▼
                    ┌───────────────┐
                    │  Mattermost   │
                    │  (4 channels) │
                    └───────────────┘

Goodhart’s Law Detection

The system calculates metrics to detect when agents are gaming for agreement rather than genuine consensus:

Consensus Quality Metrics:

• Position Diversity: How different are initial positions?
• Argument Depth: Are agents engaging with specifics?
• Concession Authenticity: Do position changes cite evidence?
• Dissent Preservation: Are minority views maintained?

When metrics drop below thresholds, the system flags potential Goodhart gaming and can inject adversarial prompts to restore genuine debate.

Vector Memory

PostgreSQL with pgvector stores debate outcomes for cross-topic learning:

• 1536-dimension embeddings via OpenAI
• HNSW indexing for fast similarity search
• RAG recall brings relevant past debates into context
• Agents learn from previous research sessions

Human-in-the-Loop

Critical decision points require human approval:

1. Research Direction: Approve debate topic framing
2. Implementation Plan: Sign off before code generation
3. Build Artifacts: Review scaffolded repositories

Signals can timeout after 7 days with configurable defaults.

Current Status

Summary