Agora

Core Problem

• In organizational decisions, minority perspectives (junior devs, designers) get drowned by senior voices
• Single LLM = single model bias dominating conclusions
• Existing approaches (Karpathy's LLM Council, Mixture-of-Agents) use majority voting/consensus → minority opinions structurally disappear

Solution — Model × Role Matrix

Each cell represents one debate participant. Assigning the same role on different models reveals model bias.

            | Claude | Gemini | Codex | GPT |
------------|--------|--------|-------|-----|
CTO         |   ✓    |        |       |     |
Junior Dev  |        |   ✓    |       |     |
PM          |        |        |  ✓    |     |
Designer    |   ✓    |        |       |     |
Executive   |        |   ✓    |       |     |

Key Features

1. Moderator-Secretary Separation

The Moderator manages debate flow, while the Secretary records proceedings in real-time with minority opinion highlights. Separating these roles ensures that minority viewpoints are explicitly captured rather than silently merged into consensus.

2. Jury Learning (Stanford CHI 2022)

• Explicit jury composition design — deliberate selection of diverse perspectives
• Structural guarantee for minority opinions — not just "considered" but formally preserved
• Dissenting Report — opposing arguments are documented as a first-class output alongside the majority recommendation

3. 3-Tier Graceful Degradation

1. Model fallback: Gemini Pro rate limit → Flash
2. CLI fallback: Codex 402 → skip panelist
3. Inline persona fallback: all external CLIs fail → Claude simulates personas directly

Principle: "Always produce a result with the best available resources"

User Scenarios

Feature Launch Decision

CTO says ready, Junior Dev worries about 70% test coverage, PM pushes for competitive timing. The final report includes majority opinion alongside a Dissenting Report preserving the Junior Dev's coverage concern.

Tech Stack Selection

The same role (Senior Dev) is assigned to both Claude and Gemini. Comparing their arguments reveals where model bias diverges from genuine technical reasoning.

Graceful Degradation in Action

5 panelists configured. Codex returns 402, Gemini hits a rate limit. The system auto-falls back — Gemini Pro → Flash, Codex panelist skipped — and the debate continues with 4 participants, producing a complete report.

Differentiation vs Existing Approaches