Exa如何使用LangGraph和LangSmith构建Web研究多智能体系统

Exa, known for their high-quality search API, recently launched their most ambitious product yet: a deep research agent that can autonomously explore the web until it finds the structured information users need.

Exa 以其高质量的搜索 API 而闻名，最近推出了他们迄今为止最雄心勃勃的产品：一个深度研究代理，可以自主探索网络，直到找到用户所需的结构化信息。

This case study explores how Exa's engineering team leveraged LangGraph to build a production-ready multi-agent system that processes hundreds of research queries daily, delivering structured results in 15 seconds to 3 minutes, depending on complexity.

本案例研究探讨了Exa的工程团队如何利用LangGraph构建一个生产就绪的多智能体系统，该系统每天处理数百个研究查询，根据复杂性在15秒到3分钟内提供结构化结果。

The evolution to agentic search

向代理搜索的演变

Exa didn't begin with agentic search, but evolved into it. The company started with a search API, then progressed to an answers endpoint that combined LLM reasoning with search results. Finally, they've now arrived at their deep research agent: their first truly agentic search API.

Exa并不是从代理搜索开始的，而是逐步演变而来的。公司最初从一个搜索API开始，然后发展到一个将LLM推理与搜索结果结合的答案端点。最终，他们现在达到了他们的深度研究代理：他们第一个真正的代理搜索API。

This reflects a broader trend across the industry: LLM applications are becoming more agentic and long-running over time. For example, we see this in research-related tasks – where what started as RAG has evolved into Deep Research. We see this in coding as well, shifting from simple auto-complete to question-answering, and now to asynchronous, long-running coding agents.

这反映了行业中的一个更广泛趋势：LLM 应用程序随着时间的推移变得更加自主和长期运行。例如，我们在与研究相关的任务中看到了这一点——最初的 RAG 已演变为深度研究。我们在编码中也看到了这一点，从简单的自动补全转变为问答，现在转向异步的长期编码代理。

This evolution is also reshaping how teams think about and utilize frameworks and tools. We've long been close partners with the Exa team via a popular open-source integration, but hadn't collaborated with them on a product until now. Their original answers endpoint didn't rely on framework, but as they transitioned to a more complex deep-research architecture, they reevaluated their options and chose to use LangGraph. This again mimics a common trend we see — as architectures get more complex,...