Back office ops · Production

Exa builds production multi-agent deep research agent with LangGraph

The problem

As Exa's architecture evolved from a simple answers endpoint to a more complex agentic deep research system, they needed a coordination framework — their original answers endpoint required no framework but could not support the new multi-agent complexity.

Workflow diagram · grounded in source
1
User submits research query
trigger
“a deep research agent that can autonomously explore the web until it finds the structured information users need”
2
Planner generates parallel tasks
ai_action
“Planner: Analyzes the research query and dynamically generates multiple parallel tasks”
3
Tasks execute with specialized tools
ai_action
“Each task receives: - Specific task instructions - A required output format (always JSON schema) - Access to specialized Exa API tools”
4
Snippet vs full content decision
validation
“Rather than automatically crawling full page content, the system first attempts reasoning on search snippets”
5
Observer aggregates context
ai_action
“While the observer maintains full visibility across all components, individual tasks only receive the final cleaned outputs from other tasks, not intermediate reasoning states”
6
Structured JSON output delivered
output
“Exa's agent maintains structured JSON output at every level. The output format can be specified at runtime”
Reported outcome

Exa built a production-ready multi-agent deep research system that processes hundreds of research queries daily, delivering structured results in 15 seconds to 3 minutes depending on complexity, with LangSmith observability informing production pricing models.

Reported metrics
Research queries processed per dayhundreds of research queries daily
Result delivery time15 seconds to 3 minutes
Reported stack
LangGraphLangSmithExa API
Source
https://blog.langchain.com/exa/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Exa built a production-ready multi-agent deep research system that processes hundreds of research queries daily, delivering structured results in 15 seconds to 3 minutes depending on complexity, with LangSmith observa…

What tools did this team use?

LangGraph, LangSmith, Exa API.

What results were reported?

Research queries processed per day: hundreds of research queries daily; Result delivery time: 15 seconds to 3 minutes (source-reported, not independently verified).

How is this back office ops AI workflow structured?

User submits research query → Planner generates parallel tasks → Tasks execute with specialized tools → Snippet vs full content decision → Observer aggregates context → Structured JSON output delivered.

Related back office ops cases