代理工作流程

人工智能 社区

技能解读

当前原文的自动摘译,尚未人工审校。

架构原则

对所有代理使用run_in_background: true,以保持主上下文最小化 使用Task工具(切勿使用TaskOutput)以避免接收完整的代理记录 代理将输出写入.claude/cache/agents/<stage>/,以便注入到后续代理中 主要对话是纯粹的编排 - 没有繁重的工作,只有协调

1. 研究代理人

强制 NIA 作为研究层 输出:研究结果

2. 策划代理

接收:作为上下文的研究代理输出 输出:实施计划

3. 验证代理

根据研究输出审查计划:经过修订的经过验证的计划

4. 实施代理

接收:经过验证的计划 + 研究背景 TDD:首先测试失败 输出:实施 + 测试

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
人工智能
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@parcadei · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
即装即用
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Agentic Workflow Pattern

Standard multi-agent pipeline for implementation tasks.

## Architecture Principles

- Use `run_in_background: true` for all agents to keep main context minimal
- Use `Task` tool (never `TaskOutput`) to avoid receiving full agent transcripts
- Agents write outputs to `.claude/cache/agents/<stage>/` for injection into subsequent agents
- Main conversation is pure orchestration — no heavy lifting, only coordination

## Workflow Stages

### 1. Research Agent
```
Task(subagent_type="oracle", run_in_background=true, prompt="""
Query NIA Oracle (via /nia-docs skill) to verify approach and gather best practices.

Output to: .claude/cache/agents/oracle/<task>-research.md
""")
```
- Enforce NIA as the research layer
- Output: Research findings

### 2. Planning Agent
```
Task(subagent_type="plan-agent", run_in_background=true, prompt="""
Read: .claude/cache/agents/oracle/<task>-research.md
Use RP-CLI to analyze the target codebase section.
Generate implementation plan informed by research.

Output to: .claude/cache/agents/plan-agent/<task>-plan.md
""")
```
- Receives: Research agent output as context
- Output: Implementation plan

### 3. Validation Agent
```
Task(subagent_type="validate-agent", run_in_background=true, prompt="""
Read: .claude/cache/agents/plan-agent/<task>-plan.md
Read: .claude/cache/agents/oracle/<task>-research.md
Review plan against research findings and best practices.

Output to: .claude/cache/agents/validate-agent/<task>-validated.md
""")
```
- Reviews plan against research
- Output: Validated plan with amendments

### 4. Implementation Agent
```
Task(subagent_type="agentica-agent", run_in_background=true, prompt="""
Read: .claude/cache/agents/validate-agent/<task>-validated.md
Read: .claude/cache/agents/oracle/<task>-research.md

TDD approach: Write failing tests FIRST, then implement.
Run tests to verify.

Output summary to: .claude/cache/agents/implement-agent/<task>-implementation.md
""")
```
- Receives: Validated plan + research context
- **TDD**: Failing tests first
- Output: Implementation + tests

### 5. Review Agent
```
Task(subagent_type="review-agent", run_in_background=true, prompt="""
Read: .claude/cache/agents/implement-agent/<task>-implementation.md
Read: .claude/cache/agents/validate-agent/<task>-validated.md
Read: .claude/cache/agents/oracle/<task>-research.md

Cross-reference implementation against plan and research.
Run tests to confirm passing.

Output to: .claude/cache/agents/review-agent/<task>-review.md
""")
```
- Cross-references all artifacts
- Confirms tests pass
- Output: Review summary

## Agent Progress Monitoring

```bash
# Watch for system reminders:
# "Agent a42a16e progress: 6 new tools used, 88914 new tokens"

# Poll for output files:
find .claude/cache/agents -name "*.md" -mmin -5

# Check task file size growth:
wc -c /tmp/claude/.../tasks/<id>.output
```

**Stuck detection:**
1. Progress reminders stop arriving
2. Task output file size stops growing
3. Expected output file not created after reasonable time

## Directory Structure

```
.claude/cache/agents/
├── oracle/
│   └── <task>-research.md
├── plan-agent/
│   └── <task>-plan.md
├── validate-agent/
│   └── <task>-validated.md
├── implement-agent/
│   └── <task>-implementation.md
└── review-agent/
    └── <task>-review.md
```

## Key Rules

1. **Never use TaskOutput** - floods context with 70k+ token transcripts
2. **Always run_in_background=true** - isolates agent context
3. **File-based handoff** - each agent reads previous agent's output file
4. **Poll, don't block** - check file system for outputs, don't wait
5. **TDD in implementation** - failing tests first, then make them pass

## Source
- Session 2026-01-01: SDK Phase 3 implementation using this pattern

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。