council

AI Council - 多 Agent 咨询

就同一个问题获取多个 AI Agent 的观点。

前置条件

至少安装以下 CLI 之一：

Claude CLI

npm install -g @anthropic-ai/claude-cli
# Set ANTHROPIC_API_KEY

Codex CLI

npm install -g @openai/codex-cli
# Set OPENAI_API_KEY

Gemini CLI

pip install google-generativeai
# Set GEMINI_API_KEY

单个 Agent 命令

Claude

# Ask Claude
claude -p --model opus "Your question here"

# Print mode (non-interactive)
claude --print "Your question"

Codex

# Ask Codex
codex -q "Your question here"

Gemini

# Ask Gemini
gemini -m pro -o text -e "" "Your question here"

并行咨询模式

基本多 Agent 查询

# Query all agents in parallel
(
  echo "=== Claude ===" && claude --print "Your question" &
  echo "=== Gemini ===" && gemini -m pro -o text -e "" "Your question" &
  echo "=== Codex ===" && codex -q "Your question" &
  wait
)

结构化比较脚本

#!/bin/bash
QUESTION="$1"

# Query in parallel, save to temp files
claude --print "$QUESTION" > /tmp/claude-response.txt &
gemini -m pro -o text -e "" "$QUESTION" > /tmp/gemini-response.txt &
wait

echo "=== Claude's Response ==="
cat /tmp/claude-response.txt

echo ""
echo "=== Gemini's Response ==="
cat /tmp/gemini-response.txt

# Cleanup
rm -f /tmp/claude-response.txt /tmp/gemini-response.txt

使用场景

代码审查视角

# Get different review perspectives
CODE=$(cat mycode.ts)

claude --print "Review this code for best practices: $CODE"
gemini -m pro -o text -e "" "Review this code for security issues: $CODE"

架构决策

QUESTION="Should I use microservices or monolith for a startup MVP with 3 developers?"

# Get multiple perspectives
claude --print "$QUESTION"
gemini -m pro -o text -e "" "$QUESTION"

调试方法

ERROR="TypeError: Cannot read property 'map' of undefined"
CONTEXT="React component fetching API data"

claude --print "Debug this error: $ERROR. Context: $CONTEXT"
gemini -m pro -o text -e "" "Explain and fix: $ERROR in $CONTEXT"

寻找共识

在获取多个 Agent 的回复后，让其中一个进行综合：

# Save individual responses first, then:
gemini -m pro -o text -e "" "Here are responses from different AI agents about [topic]:

Response 1: [Claude's response]
Response 2: [Gemini's response]

Synthesize these into a consensus view, noting:
1. Where they agree
2. Where they differ
3. The most actionable recommendation"

角色预设

查询时，你可以分配角色：

软件工程委员会

claude --print "As a software architect: [question]"
gemini -m pro -o text -e "" "As a security engineer: [question]"

代码审查委员会

claude --print "Review for maintainability: [code]"
gemini -m pro -o text -e "" "Review for performance: [code]"

创造性问题解决

claude --print "Suggest conventional solutions for: [problem]"
gemini -m pro -o text -e "" "Suggest unconventional approaches for: [problem]"

辩论模式

让 Agent 相互回应：

# Round 1: Initial positions
R1_CLAUDE=$(claude --print "Argue FOR using TypeScript in a new project")
R1_GEMINI=$(gemini -m pro -o text -e "" "Argue AGAINST using TypeScript in a new project")

# Round 2: Rebuttals
R2_CLAUDE=$(claude --print "Respond to this argument against TypeScript: $R1_GEMINI")
R2_GEMINI=$(gemini -m pro -o text -e "" "Respond to this argument for TypeScript: $R1_CLAUDE")

# Synthesis
gemini -m pro -o text -e "" "Synthesize this debate:

FOR TypeScript: $R1_CLAUDE
Rebuttal: $R2_CLAUDE

AGAINST TypeScript: $R1_GEMINI
Rebuttal: $R2_GEMINI

Provide a balanced recommendation."

最佳实践

1. 使用相同问题 - 措辞一致以实现公平比较
2. 并行运行 - 使用 & 进行后台任务以提升速度
3. 捕获输出 - 保存到文件以便比较
4. 注明是哪个 Agent - 清楚地标记各回复
5. 最后综合 - 使用一个 Agent 进行总结
6. 让 Agent 匹配任务 - Claude 负责细微之处，Gemini 负责事实，Codex 负责代码