Skip to main content
W&B Fully Connected
>
Articles
>
LLM

How to build research agents with W&B Weave and Tavily

Learn how Tavily’s real-time search and Weave’s observability make AI assistants more accurate, transparent, and reliable.
Venky Yerneni
Created on August 18|Last edited on August 25
LLMs have transformed how we build AI applications, but there’s a catch. While models stay static, the world is dynamic. A research agent trained last year may not know today’s breaking news, the latest market shifts, or newly released libraries. The result? Early agents often hallucinated, returned outdated information, or misinformed their users.
That’s where Tavily and W&B Weave come in.
Tavily brings real-time web access through search, extraction, and crawling, letting your AI agents stay current.
Weave provides a complete framework for observability and iteration, helping you trace, debug, and continuously improve those agents no matter if your application is in development or production.
In this blog, we’ll show how Tavily and Weave work together to power a research agent that can retrieve live information, generate accurate insights, and provide full visibility into its reasoning.

Why static agents fall short

Conventional LLM-based agents often fall short in the following ways:
  • Lack of awareness - public relations agents with no knowledge of breaking news.
  • Stale insights - sales agents unaware of recent market shifts.
  • Outdated tools - code agents missing new libraries and docs.
In short, hallucinations, misinformation, outdated knowledge, and blind spots create significant risks to developing agents that are safe and reliable.

How Tavily fixes the problem

Tavily was built to solve this by giving agents a web-native context stack with real-time tools such as:
  • Search - find relevant, up-to-date content.
  • Extract - pull structured information from web pages.
  • Crawl - navigate and collect from multiple sources.
  • Map - build contextual knowledge for the agent.
In practice, that means your agent can now:
  • Fetch the latest financial filings for a company.
  • Grab real-time news on a market event.
  • Crawl technical docs to stay up to date with new APIs.
Minimal Tavily search example
from tavily import TavilyClient
import os

def tavily_search(query: str) -> dict:
    """
    Search using Tavily API and return the response.
    """
    tv = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
    return tv.search(query)

# Example
if __name__ == "__main__":
    print(tavily_search("Who is Leo Messi?"))

How W&B Weave makes it trustworthy

But retrieval alone isn’t enough when building agents, you also need to understand how your agent reached its conclusions. That’s where W&B Weave comes in.
With Weave, you can:
  • Trace agent workflows - every Tavily API call is logged.
  • Debug errors - spot broken API calls or failed retrievals.
  • Measure performance - track latency, token usage, and results.
  • Experiment safely - iterate on prompts, models, and guardrails.
For example, wrapping your Tavily calls with @weave.op() ensures every step is captured and visualized in Weave’s UI. You can drill down into the trace tree, inspect inputs/outputs, and compare different execution paths.
Add Weave in just 3 lines (import, init & decorate)
import weave                             # (1) Import weave
from tavily import TavilyClient
import os

weave.init(project="your-project-name")  # (2) connect to your Weave project

@weave.op()                              # (3) trace this function
def tavily_search(query: str) -> dict:    
    tv = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
    return tv.search(query)

# Example
if __name__ == "__main__":
    print(tavily_search("Who is Leo Messi?"))

Demo: The research agent in action

Let’s walk through an example to see how our research agent uses Tavily’s real-time tools in combination with Weave’s observability.

Agent workflow visualization

The research process unfolds with Tavily

Smart summarization at work

Weave observability in action


This demo showcases how Tavily's comprehensive toolkit (search, extract, crawl) combined with Weave's observability creates a production-ready research agent that's both powerful and transparent.
Try out our full example here: https://github.com/tavily-ai/tavily-chat/tree/feat/weave

Conclusion

Building intelligent research agents doesn’t have to mean sacrificing accuracy or trust. Tavily gives your agents access to the live web through search, crawl, and extract. Weave ensures you can trace, debug, and improve those agents over time. Together, they turn static, unreliable assistants into context-aware, trustworthy research agents.
By combining Tavily’s real-time retrieval with Weave’s observability, you get the best of both worlds: assistants that are always up to date and fully transparent in how they operate. Whether you are building tools for financial analysis, technical research, or customer support, this stack ensures your agents are fast, grounded, and easy to iterate on.
This partnership is about more than just solving today’s limitations. It is about creating the foundation for the next generation of AI applications, where every answer is both reliable and explainable.
Join us for our upcoming webinar on Tuesday, August 26th to see Tavily and Weave in action. We will walk through building the research agent step by step and share best practices for scaling your own AI assistants. RSVP and reserve your spot today.
Tags: Articles, LLM

Created with ❤️ on Weights & Biases.

https://wandb.ai/venky17/tavily-report/reports/How-to-build-research-agents-with-W-B-Weave-and-Tavily--VmlldzoxNDAzMDY5NQ
Made with Weights & Biases. Sign up or log in to create reports like this one.
Iterate on AI agents and models faster. Try Weights & Biases today.
Sign upTry W&B now