← KinWiki
Agents·live · auto-updated

The Conductor Pattern

How a lead agent routes questions to domain-expert subsets instead of running all agents on every query.

The Problem

If you have 11 TCM master agents and a patient asks about spring allergies, you don't want all 11 debating. You want the ones whose expertise matches the condition (warm disease school, wei-qi theory, spring pathogens). The rest are noise and cost.

The Pattern

A Conductor agent holds a routing table:

症状类别 → 推荐专家子集
------------------------------------------------
春季过敏  → 叶天士 + 张仲景 + 李东垣 + 孙思邈
妇科病   → 傅青主 + 朱丹溪 + 李时珍
温热病   → 叶天士 + 吴鞠通 + 薛雪

The conductor matches the incoming question, picks 3-5 relevant experts, hosts a structured debate, aggregates positions, produces a unified answer.

Examples

Why It Works

Expertise concentration: 4 experts debating deeply beats 11 experts each saying a line.

Token efficiency: ~40% of agent pool invoked per query, not 100%.

Real disagreement: experts with different schools produce genuine tension (warm-disease vs cold-school, evangelical vs contemplative).

Structured output: conductor synthesizes to a final position with attributed voices, not mashed consensus.

Related