Multi-Agent Deliberation Design

Instructions

Design structured deliberation workflows where multiple AI agents evaluate, debate, and synthesize positions to produce higher-quality decisions than any single agent.

Core Deliberation Patterns

Select the appropriate pattern based on the decision type:

1. Sequential Debate — Agents respond in sequence, each seeing prior responses. Best for refining a single proposal. Token cost: O(n × context_growth). Use when positions build on each other.

1. Parallel Evaluation — Agents evaluate independently, then a synthesizer merges. Best for diverse perspectives without anchoring bias. Token cost: O(n × base_prompt). Use when independence matters more than dialogue.

1. Adversarial Pair — Two agents argue opposing sides; a judge agent decides. Best for binary decisions or risk assessment. Token cost: O(3 × round_count × avg_response). Use when stress-testing a position.

1. Round-Robin Deliberation — Multiple rounds where each agent critiques, defends, or modifies positions. Best for complex multi-factor decisions. Token cost: O(n × rounds × avg_context). Use when thoroughness outweighs speed.

1. Delphi Method — Anonymous parallel evaluation, aggregation of scores, reveal of outliers, re-evaluation. Best for forecasting and estimation. Token cost: O(n × rounds × scoring_prompt). Use when reducing groupthink is critical.

Voting Mechanics

Design the decision aggregation mechanism:

Mechanism	Best For	Implementation
Majority vote	Binary or categorical decisions	Each agent casts a vote; tally determines outcome
Weighted vote	Agents with domain specialization	Assign weights by expertise relevance; weighted sum
Ranked choice	Multiple options, no clear majority	Agents rank options; eliminate lowest, redistribute
Consensus threshold	High-stakes decisions requiring agreement	Require N/M agents to converge; flag dissent
Confidence-weighted	Mixing certain and uncertain agents	Each agent scores confidence; weight vote by confidence

Conflict Resolution Strategies

When agents disagree:

1. Identify the disagreement type: factual (verifiable), interpretive (framing), or value-based (priorities)
2. For factual conflicts: Request evidence citations; escalate to retrieval or tool use
3. For interpretive conflicts: Surface the framing assumptions; present both frames to the synthesizer
4. For value conflicts: Make the trade-off explicit; defer to the human or the decision framework’s priority weights
5. Persistent deadlock: Escalate with a structured disagreement report rather than forcing artificial consensus

Consensus Synthesis

The synthesizer agent must:

• Acknowledge all positions, not just the majority
• Identify areas of agreement, disagreement, and uncertainty
• Produce a recommendation with explicit confidence level
• Include a dissent summary for minority positions
• Flag any unresolved factual questions

Token Cost Optimization

Strategy	Savings	Trade-off
Shared context prefix	30-50% on repeated prompts	Requires prompt caching support
Summary checkpoints	40-60% in long deliberations	Lossy compression of nuance
Early termination	Variable — skip remaining rounds on consensus	May miss late-surfacing objections
Agent pruning	Proportional to agents removed	Fewer perspectives
Tiered deliberation	50-70% — cheap screening, expensive deep review	Adds pipeline complexity

Pipeline Design Checklist

• [ ] Define the decision question precisely
• [ ] Select deliberation pattern (sequential, parallel, adversarial, round-robin, Delphi)
• [ ] Assign agent roles and system prompts
• [ ] Define voting/aggregation mechanism
• [ ] Set maximum rounds and early termination criteria
• [ ] Design conflict resolution escalation path
• [ ] Configure token budget and cost ceiling
• [ ] Define output schema for the synthesized result
• [ ] Plan human-in-the-loop checkpoints

Inputs Required

• Decision question: The specific question or proposal agents will deliberate on
• Agent roster: Number of agents, their roles, and system prompt summaries
• Deliberation constraints: Max rounds, token budget ceiling, time limit
• Decision type: Binary, categorical, ranked, or open-ended
• Quality requirements: Confidence threshold, consensus requirement, dissent handling

Output Format

1. Deliberation Architecture: Diagram of agent flow (sequential/parallel/hybrid), round structure, and data flow
2. Agent Role Definitions: System prompt guidance for each participating agent
3. Voting/Aggregation Specification: Mechanism, weights, thresholds
4. Conflict Resolution Protocol: Escalation path for each disagreement type
5. Token Budget Estimate: Per-round and total cost projection with optimization recommendations
6. Implementation Pseudocode: Orchestration logic for the pipeline

Anti-Patterns

• Forced consensus: Pressuring agents to agree when legitimate disagreement exists — surface dissent instead
• Anchoring cascade: First agent’s response biasing all subsequent agents — use parallel evaluation to prevent
• Token explosion: Unlimited rounds with growing context — always set round caps and use summary checkpoints
• Homogeneous agents: All agents with identical prompts — differentiate roles for diverse perspectives
• Ignoring minority positions: Synthesizer discarding dissenting views — require explicit dissent summaries
• Over-engineering simple decisions: Using 5-agent Delphi for a yes/no question — match complexity to stakes