English
How we built our multi-agent research system
Source: Anthropic Engineering Blog URL: https://www.anthropic.com/engineering/multi-agent-research-system Published: Jun 13, 2025
Architecture: Orchestrator-Worker Pattern
Lead agent plans research process, spawns subagents that explore different aspects in parallel. Subagents operate with their own context windows — enables compression of findings for the lead agent.
Key Benefits
- Multi-agent outperformed single-agent by 90.2% on internal research eval (Opus 4 lead + Sonnet 4 subagents)
- Token usage explains 80% of variance in BrowseComp evaluation
- Parallel tool calling cut research time by up to 90%
Prompt Engineering Lessons
1. Think like your agents — build simulations with exact prompts/tools, watch step-by-step 2. Teach the orchestrator how to delegate — objective, output format, tools/sources guidance, task boundaries 3. Scale effort to query complexity — simple: 1 agent + 3-10 calls; comparison: 2-4 agents + 10-15 calls; complex: 10+ agents with clear divisions 4. Tool design is critical — bad tool descriptions send agents down wrong paths; each tool needs distinct purpose and clear description 5. Let agents improve themselves — Claude 4 models can be excellent prompt engineers; testing agent improved MCP tool descriptions → 40% decrease in task completion time 6. Start wide, then narrow — agents default to overly specific queries; counter by prompting for short broad queries first, then progressively narrow 7. Guide the thinking process — extended thinking as controllable scratchpad for planning
Why Multi-Agent Works
Three factors explain 95% of performance variance in browsing agent evaluations: 1. Token usage (80% of variance) 2. Number of tool calls 3. Model choice
Multi-agent architectures effectively scale token usage for tasks that exceed single agent limits. Latest Claude models act as large efficiency multipliers on token use.
Tradeoffs
- Agents use ~4× more tokens than chats
- Multi-agent systems use ~15× more tokens than chats
- Not suited for tasks requiring shared context or many dependencies between agents
- Best for high-value tasks with heavy parallelization, information exceeding single context windows, and many complex tools
中文
我们如何构建多 Agent 研究系统
来源： Anthropic 工程博客 链接： https://www.anthropic.com/engineering/multi-agent-research-system 发布日期： 2025年6月13日
架构：编排器-工作者模式
主 agent 规划研究过程，生成在多个方向并行探索的子 agent。子 agent 拥有自己的上下文窗口 — 允许对发现进行压缩后汇报给主 agent。
主要优势
- 多 Agent 在内部研究评估中超出单 Agent 90.2%（Opus 4 主 + Sonnet 4 子）
- Token 使用解释了 BrowseComp 评估中 80% 的方差
- 并行工具调用将研究时间缩短多达 90%
提示工程经验
1. 像 agent 那样思考 — 用精确的提示/工具构建模拟，一步步观察 2. 教编排器如何委托 — 目标、输出格式、工具/来源指导、任务边界 3. 根据查询复杂度调整投入 — 简单：1 agent + 3-10 次调用；比较：2-4 agents + 10-15 次调用；复杂：10+ agents 且分工明确 4. 工具设计至关重要 — 糟糕的工具描述会让 agent 走错路；每个工具需要明确用途和清晰描述 5. 让 agent 自我改进 — Claude 4 模型可以成为出色的提示工程师；测试 agent 改进了 MCP 工具描述 → 任务完成时间减少 40% 6. 先宽后窄 — agent 默认使用过于具体的查询；通过提示先做简短宽泛查询，然后逐步收窄来对抗 7. 引导思维过程 — 扩展思考作为可控制的草稿纸用于规划
多 Agent 为何有效
浏览 agent 评估中 95% 的性能方差来自三个因素： 1. Token 使用（80% 的方差） 2. 工具调用次数 3. 模型选择
多 Agent 架构有效地为超出单个 agent 限制的任务扩展 token 使用。最新 Claude 模型在 token 使用上充当大型效率倍增器。
权衡
- Agent 比聊天多使用约 4 倍的 token
- 多 Agent 系统比聊天多使用约 15 倍的 token
- 不适用于需要共享上下文或 agent 之间有很多依赖的任务
- 最适合高价值任务：高度并行化、超过单个上下文窗口的信息量、众多复杂工具