42 lines
1.4 KiB
Markdown
42 lines
1.4 KiB
Markdown
---
|
||
title: "Context Passing"
|
||
type: concept
|
||
tags: [multi-agent, workflow, communication]
|
||
last_updated: 2026-04-21
|
||
---
|
||
|
||
## Definition
|
||
上下文传递（Context Passing）是一种多智能体通信模式，在智能体之间传递完整的上下文信息，而非摘要或压缩版本。
|
||
|
||
## Core Principle
|
||
**"Copy-paste agent outputs between steps — don't summarize, use the full output"**
|
||
|
||
## Problem It Solves
|
||
- 摘要会丢失关键细节
|
||
- 压缩可能遗漏重要上下文
|
||
- 智能体不具备共享记忆
|
||
|
||
## Best Practices
|
||
1. **完整复制**：将上一智能体的输出完整粘贴到下一智能体的输入
|
||
2. **不要摘要**：即使很长也要完整传递
|
||
3. **不要解释**：让接收智能体自己解析和处理完整信息
|
||
4. **包含所有内容**：原始输出、元数据、证据、决策理由等
|
||
|
||
## Example
|
||
在 Multi-Agent Workflow: Startup MVP 中：
|
||
```
|
||
Sprint Prioritizer 输出 → [完整粘贴] → Backend Architect 输入
|
||
UX Researcher 输出 → [完整粘贴] → Backend Architect 输入
|
||
Backend Architect 输出 → [完整粘贴] → Frontend Developer 输入
|
||
```
|
||
|
||
## Trade-offs
|
||
- **优点**：最大化信息保留，便于接收智能体做出完整判断
|
||
- **缺点**：消耗更多上下文窗口，可能需要更长的处理时间
|
||
|
||
## Related Concepts
|
||
- [[Sequential Handoffs]]：顺序交接模式
|
||
- [[Parallel Work]]：并行工作模式
|
||
- [[Quality Gates]]：质量门控
|
||
- [[Multi-Agent Team]]：多智能体团队
|