northseadl/norix-skills / llm-agent-dev
llm-agent-dev
UnverifiedLLM Agent engineering: pattern selection (12-mode matrix), data simulation, convergence iteration. Covers intent routing, function calling, ReAct, MCP, prompt chaining, guardrails, evaluation.
references/patterns/08-handoff-delegation.md
View on GitHub# 模式 ⑧ Handoff / Agent 委托
> Agent 间的任务委托和转交,支持专业化分工和人工升级
> 延迟取决于委托链 · 可预测性 ★★★☆☆ · OpenAI Agents SDK 核心原语
## 背景
Handoff 是 OpenAI Agents SDK (Swarm 的生产版) 的核心原语之一。
它允许一个 Agent 将对话或任务"转交"给另一个更专业的 Agent,
或升级到人工处理。这是构建生产级多 Agent 系统的基础模式。
## 适用场景
- 客服系统中不同专业领域的转接(销售→技术→退款)
- Agent 能力边界明确,需要委托给专家
- 人工升级(HITL)作为 Handoff 的特例
- 渐进式复杂度:简单 Agent 处理不了时委托给复杂 Agent
## 核心概念
```typescript
// Handoff 定义
interface Handoff {
target: Agent; // 目标 Agent
condition?: string; // 触发条件
contextTransfer: 'full' | 'summary' | 'selective';
onHandoff?: (context: AgentContext) => void;
}
// Agent 可以声明自己的 handoff 列表
interface Agent {
id: string;
name: string;
instructions: string;
tools: ToolDefinition[];
handoffs: Handoff[]; // 可委托给哪些 Agent
}
```
## 实现
### Agent-to-Agent Handoff
```typescript
class HandoffableAgent implements SkillHandler {
id: string;
name: string;
private llm: LLMClient;
private tools: ToolDefinition[];
private handoffs: Map<string, HandoffableAgent>;
async execute(params: Record<string, any>, context: AgentContext): Promise<SkillResponse> {
const response = await this.llm.chat({
model: 'haiku',
messages: this.buildMessages(params, context),
tools: [
...this.tools,
// 将 handoff 目标暴露为特殊 "tool"
...this.handoffsAsTools(),
],
});
// 检查是否触发 handoff
if (response.tool_calls?.some(tc => tc.function.name.startsWith('handoff_to_'))) {
const handoffCall = response.tool_calls.find(tc => tc.function.name.startsWith('handoff_to_'))!;
const targetId = handoffCall.function.name.replace('handoff_to_', '');
const target = this.handoffs.get(targetId);
if (target) {
// 转移上下文并委托
const transferredContext = this.transferContext(context, handoffCall);
return target.execute(params, transferredContext);
}
}
return this.formatResponse(response);
}
private handoffsAsTools(): ToolDefinition[] {
return [...this.handoffs.entries()].map(([id, agent]) => ({
name: `handoff_to_${id}`,
description: `将对话转交给 ${agent.name}。当前对话超出你的能力范围时使用。`,
parameters: {
type: 'object',
properties: {
reason: { type: 'string', description: '转交原因' },
summary: { type: 'string', description: '对话摘要' },
},
required: ['reason'],
},
}));
}
}
```
### Agent-to-Human Handoff (HITL)
```typescript
const humanHandoff: Handoff = {
target: humanAgent,
condition: 'confidence < 0.5 || sentiment === "negative" || topic === "complaint"',
contextTransfer: 'full',
onHandoff: async (context) => {
// 通知人工客服
await notifyHumanAgent({
sessionId: context.sessionId,
userId: context.userId,
summary: context.conversationSummary,
priority: context.sentiment === 'negative' ? 'high' : 'normal',
});
},
};
```
## 委托拓扑
### 线性委托链
```
用户 → 前台 Agent → 专家 Agent → 执行完返回
```
### 星型委托
```
┌→ 销售 Agent
用户 → 路由 Agent ─→ 技术 Agent
└→ 退款 Agent
```
### 渐进式升级
```
用户 → 简单 Agent (CREX)
│ 处理不了
▼
中级 Agent (FC + Slot)
│ 仍处理不了
▼
高级 Agent (ReAct)
│ 仍处理不了
▼
人工客服 (HITL)
```
## 与 Router-Expert 的区别
| 维度 | Router-Expert (⑥) | Handoff (⑧) |
|------|-------------------|-------------|
| 决策者 | 路由层(外部分类器) | Agent 自身(LLM 判断) |
| 时机 | 对话开始时路由 | 对话中间随时转交 |
| 可逆性 | 不可逆(选定专家后固定) | 可逆(目标可以再 handoff 回来) |
| 上下文 | 分割 | 可选择性转移 |
## ACT 协议 (Agentic Commerce Trust)
阿里 2026.01 推出的 ACT 协议是 Handoff 在电商场景的标准化:
```
用户: "帮我点个外卖"
→ 通义千问 Agent
→ Handoff to 饿了么 MCP Server (餐厅发现+菜品)
→ Handoff to 支付宝 Agent (优惠券+支付)
→ 结果聚合返回用户
```
Agent-to-Agent 之间通过标准协议建立信任,互相委托子任务。
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |