我们的研究功能使用多个 Claude 智能体更有效地探索复杂主题。我们分享了构建该系统时遇到的工程挑战和学到的经验。Claude 现在具备研究能力，可以搜索网络、Google Workspace 以及任何集成，以完成复杂任务。这个多智能体系统从原型到生产的历程，让我们在系统架构、工具设计和提示工程方面学到了关键经验。多智能体系统由多个智能体（LLM 在循环中自主使用工具）协同工作组成。我们的研究功能涉及一个智能体，它根据用户查询规划研究过程，然后使用工具创建并行搜索信息的智能体。具有多个智能体的系统在智能体协调、评估和可靠性方面带来了新的挑战。
Our Research feature uses multiple Claude agents to explore complex topics more effectively. We share the engineering challenges and the lessons we learned from building this system. Claude now has Research capabilities that allow it to search across the web, Google Workspace, and any integrations to accomplish complex tasks. The journey of this multi-agent system from prototype to production taught us critical lessons about system architecture, tool design, and prompt engineering. A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together. Our Research feature involves an agent that plans a research process based on user queries, and then uses tools to create parallel agents that search for information simultaneously. Systems with multiple agents introduce new challenges in agent coordination, evaluation, and reliability.
核心贡献 · Key contributions
采用编排器-工作器模式的多智能体研究系统在内部评估中比单智能体性能高出 90.2%。 Multi-agent research system using orchestrator-worker pattern outperforms single-agent by 90.2% on internal eval.
具有独立上下文窗口的并行子智能体实现同时探索，将研究时间减少高达 90%。 Parallel subagents with separate context windows enable simultaneous exploration, reducing research time by up to 90%.
提示工程原则：像智能体一样思考、教授委派、按查询复杂度分配资源、设计工具、引导思考过程。 Prompt engineering principles: think like agents, teach delegation, scale effort, design tools, and guide thinking process.
使用评分标准（0.0-1.0）和及格/不及格等级的 LLM 作为评判者的评估方法，可有效扩展至自由形式的研究输出。 LLM-as-judge evaluation with rubric scores (0.0-1.0) and pass-fail grade scales effectively for free-form research outputs.
多智能体系统擅长需要大量并行化的广度优先查询，超出单一上下文窗口的限制。 Multi-agent systems excel at breadth-first queries with heavy parallelization, exceeding single context windows.
局限 · Limitations
多智能体系统使用的 Token 量约为聊天的 15 倍，经济可行性仅限于高价值任务。 Multi-agent systems use about 15× more tokens than chats, limiting economic viability to high-value tasks.
智能体间依赖关系多的编码任务不适合，因为实时协调能力有限。 Coding tasks with many dependencies between agents are not a good fit due to limited real-time coordination.
同步执行造成瓶颈；主导智能体无法引导子智能体，需等待所有子智能体完成。 Synchronous execution creates bottlenecks; lead agent cannot steer subagents and waits for all to complete.
微小提示变化引发的涌现行为可能不可预测地改变子智能体行为，增加调试难度。 Emergent behaviors from small prompt changes can unpredictably alter subagent behavior, complicating debugging.
部署需要谨慎协调，采用彩虹部署以避免中断运行中的智能体。 Deployment requires careful coordination with rainbow deployments to avoid breaking running agents mid-process.
论文章节 · Sections（共 8）
概述Overview
多智能体系统的优势Benefits of a multi-agent system
研究系统架构概述Architecture overview for Research
研究型智能体的提示工程与评估Prompt engineering and evaluations for research agents
智能体的有效评估Effective evaluation of agents
生产可靠性与工程挑战Production reliability and engineering challenges