The Deliberation Model

Council’s core value proposition is structured deliberation — not just “ask multiple experts,” but coordinated, disagreement-aware debate between distinct perspectives that synthesize into actionable insights.

The Problem with Single-Agent Advice

A single LLM — even a frontier model — provides one perspective compressed into a “balanced” answer. It optimizes for agreement with itself, not for surfacing tensions between competing priorities. You get generic pros/cons lists instead of stakeholder-specific reasoning.

Council’s Approach: Multi-Agent Debate

Council runs multiple expert agents in parallel, each with:

• A distinct epistemic stance (how they form beliefs)
• Weighted evidence types (what they prioritize)
• Disagreement incentives (enforced through anti-sycophancy)
• Persistent memory (they remember past debates)

The orchestrator ensures experts see each other’s turns, prompts them to find weaknesses in prior arguments, and synthesizes the conversation into decision-ready output.

Two Debate Modes

Freeform Mode (default)

Rounds of expert turns controlled by a moderator strategy:

• Round-robin: everyone speaks once per round, sequentially
• Devil’s advocate: one expert is prompted to challenge every claim
• Consensus-check: terminates early when the panel converges

The moderator decides turn order, who speaks next, and what prompt they receive. Flexible, but can meander without strong moderation.

Structured Mode

Fixed 4-phase choreography (inspired by structured debate formats):

1. Opening — each expert states their position and reasoning
2. Cross-examination — experts critique each other’s claims
3. Rebuttal — experts defend or adjust their stance
4. Synthesis — the orchestrator summarizes convergence, tensions, and recommendations

Structured mode guarantees depth and coverage, but takes more tokens and time.

Why This Works

Multi-agent debate surfaces:

• Trade-offs: CTO argues scalability, CFO argues cost — the tension is the insight
• Blind spots: Security expert finds what the feature team missed
• Confidence signals: when two domain experts disagree, that’s a risk flag

The synthesis step converts the debate into a decision artifact: what the panel agrees on, where they diverge, which tensions are unresolvable, and what data would close the gap.

Context Window Management

Long debates hit token limits. Council offers:

• Visibility scoping: show experts only their round (same-round), recent turns (recent), or everything (all)
• Rolling summaries: after N rounds, replace old verbatim turns with an LLM-generated summary
• Hard caps: maxPromptChars truncates context to fit models with smaller windows

These controls trade memory for affordability — a 10-round architecture debate with full context might cost $5; with summaries-after-3 and same-round visibility, under $1.

Relation to Other Concepts

• Anti-Sycophancy — how Council enforces disagreement
• Moderation Strategies — how turn order is decided
• Context Management — how long debates stay within token budgets
• Memory Model — how experts remember past debates