Exa Innovates with Multi-Agent Web Research System Using LangGraph

Exa, a prominent player in the search API industry, has unveiled its latest innovation: a sophisticated multi-agent web research system. This development is powered by LangGraph and LangSmith, and it aims to revolutionize how complex research queries are processed, according to LangChain.

The Evolution to Agentic Search

Exa’s journey to this advanced system began with a simple search API. Over time, the company evolved their offerings to include an answers endpoint that integrated large language model (LLM) reasoning with search results. The latest development is their deep research agent, marking their entry into truly agentic search APIs. This reflects a broader industry trend towards more autonomous and long-running LLM applications.

The transition to a deep-research architecture prompted Exa to adopt LangGraph, which has become a preferred framework for handling increasingly complex architectures. This shift aligns with industry movements where simpler setups are upgraded to handle more sophisticated tasks, such as research and coding.

Designing a Multi-Agent System

Exa’s system features a multi-agent architecture built on LangGraph, consisting of:

1. Planner: Analyzes queries and generates parallel tasks.
2. Tasks: Executes independent research using specialized tools.
3. Observer: Oversees the entire process, maintaining context and citations.

This architecture allows dynamic scaling, adjusting the number of tasks based on the query’s complexity. Each task is provided with specific instructions, required output formats, and access to Exa’s API tools, ensuring efficient processing from simple to complex queries.

Key Design Insights

Exa’s system emphasizes structured output and efficient resource usage. By prioritizing reasoning on search snippets before full content retrieval, the system reduces token usage while maintaining research quality. This approach is vital for API consumption, where reliable and structured JSON outputs are crucial.

Exa’s design choices draw inspiration from other industry leaders, such as the Anthropic Deep Research system, incorporating best practices in context engineering and structured data output.

Utilizing LangSmith for Observability

LangSmith’s observability features, particularly in token usage tracking, played a critical role in Exa’s system development. This capability provided essential insights into resource consumption, informing pricing models and optimizing performance.

Mark Pekala, a software engineer at Exa, emphasized the importance of LangSmith’s ease of setup and its contribution to understanding token usage, which was pivotal for the system’s cost-effective scalability.

Conclusion

Exa’s innovative use of LangGraph and LangSmith showcases the potential of multi-agent systems in handling complex web research queries efficiently. The project highlights key takeaways for similar endeavors, such as the importance of observability, reusability, structured outputs, and dynamic task generation.

As Exa continues to refine its deep research agent, this development serves as a model for building robust, production-ready agentic systems that deliver substantial business value.

Image source: Shutterstock