feat(wiki): ingest misc sources batch (Daily notes/Others/Skills/Vibe Coding/微信公众号/跨境电商)
This commit is contained in:
42
wiki/concepts/多Agent可靠性模式.md
Normal file
42
wiki/concepts/多Agent可靠性模式.md
Normal file
@@ -0,0 +1,42 @@
|
||||
---
|
||||
title: "多Agent可靠性模式"
|
||||
type: concept
|
||||
tags: []
|
||||
---
|
||||
|
||||
## Definition
|
||||
多智能体系统中用于克服 LLM 不可靠性(幻觉、逻辑谬误、上下文漂移)的四大架构模式:层级结构、共识、 adversarial debate 和淘汰制。
|
||||
|
||||
## Four Patterns
|
||||
|
||||
### 1. 层级结构(Hierarchy)
|
||||
- **角色**:Planner(规划器)+ Worker(工作者)+ Validator(验证器)
|
||||
- **依赖图强制协作**:Worker 必须等待 Planner 分配任务,且无法作弊(Validator 会发现)
|
||||
- **适用场景**:需要将上下文分开的复杂工作流程
|
||||
|
||||
### 2. 共识(Consensus / Voting)
|
||||
- **机制**:N 个 LLM 独立生成同一任务答案,取多数票
|
||||
- **数学基础**:3 个模型同时产生相同幻觉的概率 = 0.2³ = 0.8%(假设单模型幻觉率 20%)
|
||||
- **适用场景**:事实核查、分类任务
|
||||
|
||||
### 3. 对抗辩论(Adversarial Debate)
|
||||
- **角色**:Generator → Critic(反对) → Judge(裁决)
|
||||
- **机制**:Truth survives the fight,真理越辩越明
|
||||
- **适用场景**:安全分析、代码审查、高风险内容审核
|
||||
|
||||
### 4. 淘汰制(Knock-out)
|
||||
- **类比**:SRE 中服务器是"cattle"(可替换)而非"pets"(独一无二)
|
||||
- **机制**:N 个 Agent 执行任务,最差者被淘汰;可选择用获胜者特征替换已淘汰者
|
||||
- **适用场景**:迭代式 Agent 工程、开发调试
|
||||
|
||||
## Core Insight
|
||||
> "Stop treating LLMs like magic chatbots. Start treating them like unreliable components in a distributed system."
|
||||
|
||||
## Related Concepts
|
||||
- [[LLM不可靠性]]
|
||||
- [[验证器模式]]
|
||||
- [[共识投票]]
|
||||
- [[AI拟人化谬误]]
|
||||
|
||||
## Sources
|
||||
- [[Multi-Agent-System-Reliability.md]]
|
||||
Reference in New Issue
Block a user