Building a Deep Research Agent powered by Tavily and Token Factory

A research agent plans, searches, reads, revises, and cites where each claim came from, instead of answering from training data alone. This cookbook wires Tavily web search into a Deep Agent driven by a Nebius Token Factory model, so the answer is grounded in pages fetched at runtime.

What you will buildWhat you will build

A research agent that takes a question, decomposes it into sub-questions, runs advanced Tavily searches, and returns a structured report. It has three pieces:

• MiniMax-M3 on Nebius Token Factory as the reasoning model, reached through langchain-nebius.
• Tavily Search as the agent’s tool, configured for advanced depth and five results per query.
• deepagents to supply the planning loop, the virtual file system for notes, and the sub-agents that turn a chat model into a researcher.

A measured run of the default question made 6 model requests, cost about $0.07 at Token Factory pricing, and produced a 20,000-character report in under 90 seconds.

PrerequisitesPrerequisites

• Python 3.12 or newer
• A Nebius Token Factory account with access to MiniMaxAI/MiniMax-M3
• A Tavily account and API key
• Both credentials in environment variables, never in a notebook cell or a commit

Run the cookbookRun the cookbook

1. Install the dependencies:

   python -m venv .venv && source .venv/bin/activate
   pip install -U deepagents langchain langchain-core langchain-nebius langchain-tavily tavily-python
   

2. Export both credentials:

   export NEBIUS_API_KEY="your-token-factory-key"
   export TAVILY_API_KEY="your-tavily-key"
   

3. Build the agent from a search tool, a model, and one call that combines them:

   from deepagents import create_deep_agent
   from langchain_nebius import ChatNebius
   from langchain_tavily import TavilySearch
   
   tavily_search = TavilySearch(max_results=5, search_depth="advanced")
   model = ChatNebius(model="MiniMaxAI/MiniMax-M3")
   agent = create_deep_agent(model=model, tools=[tavily_search])
   

4. Ask a research question and print the report:

   result = agent.invoke({
       "messages": [
           {
               "role": "user",
               "content": "Research GPUs available in the US in 2026 and write a detailed report.",
           }
       ]
   })
   
   print(result["messages"][-1].content)
   

Everything between the question and the report is the agent deciding what to search for, what to read, and when it has enough to write.

Verify the resultVerify the result

A successful run returns a report with concrete, checkable specifics (named products, dates, figures) and the sources Tavily surfaced. If the output reads like a generic summary with no sources, the search tool was never invoked; check that TAVILY_API_KEY is set in the same environment.

To see the agent’s planning steps and tool calls rather than just the answer, print the whole state instead of the last message:

import json
print(json.dumps(result, indent=2, default=str))

TroubleshootingTroubleshooting

• 401 from Token Factory: the key is missing, expired, or scoped to a different project. Re-export NEBIUS_API_KEY and retry.
• Empty or missing search results: confirm the Tavily key is active and your plan has remaining credits.
• The run takes a minute or two: this is expected. A research pass makes several model calls, and advanced search depth is slower than basic.

Clean upClean up

The cookbook creates no cloud resources, so there is nothing to tear down:

deactivate && rm -rf .venv
unset NEBIUS_API_KEY TAVILY_API_KEY

If you ever pasted either key into a notebook, a shell history file, or a commit, rotate it in the Token Factory console and the Tavily dashboard rather than deleting the file and assuming it is gone.