Learn How to Build Tool-Calling Agents with LangGraph

It’s 2025 and agents are right under our fingertips, but there’s a problem…

The problem—at least when it comes to building agentic systems—is that frameworks for building them, like LangGraph (an evolution of LangChain but designed specifically for building agents), are difficult to learn.

In this video I’ll show you how it works, and how you can use it to build very powerful agents (which—in this case—do very important science 🧪)

Check out the video on YouTube

AI Engineer Roadmap Launch 🚀

The AI Engineer Roadmap is live! You can sign up and start tracking your progress on https://zazencodes.com/

I’ve got a discount code that’s valid during launch month—use ZAZEN33 to get 33% off (ends March 31st 2025)

Topics from this week’s video

• Tool Demonstration

  • Code available on GitHub

  • Running the agent via Python:

    • Chose "Neutron stars" as the topic

    • Archive API fetches related papers

    • Selects a paper and downloads the PDF

    • Extracts and summarizes the paper using LLM

    • Suggests future research directions

    • Drafts a research paper

    • Renders the draft into a LaTeX PDF

    • Provides download link for the generated PDF

• Code Overview (Workflow 1)

  • Uses OpenAI model but can switch to Claude

  • Creation of a React agent (Reason + Act)

  • Tools are provided to the agent

  • Uses an infinite loop with graph.stream for continuous interaction

  • Initial prompt guides the agent behavior

• Switch to Workflow 2 (Lower-Level LangGraph Implementation)

  • Key components:

    • State management with message history

    • Definition of tools and tool nodes

    • Switching models (demonstrated with Anthropic Sonnet)

    • Binding tools to models via prompt augmentation

• Local Model Demo with Ollama

  • Switched to running a local Llama3 Grok model

  • Successfully made tool calls locally

  • Demonstrated GPU usage

  • Limitations with model context windows and performance

• Deep Dive into Tool Usage in LangGraph

  • How tools are defined using decorators (example: getWeather)

  • Use of Literal from typing to constrain tool arguments

  • Binding tools to models to include them in prompt context

  • Prompt augmentation explains tool capabilities to the LLM

  • Tools included:

    • Archive Search (fetches research papers)

    • Read PDF (extracts PDF contents)

    • Render LaTeX PDF (creates a PDF from LaTeX content)

• Closing Thoughts

  • Importance of responsible use of APIs (e.g., Archive API)

  • Results are for demonstration, not actual scientific research

  • The complexity of LangGraph and agent systems

---

For the physicist or engineer, two systems that obey the same equations have a kind of identity-or at least an analogy. And that, after all, is all our word analog means. A digital watch is nothing like the sun; an analog watch is the memory of a shadow's circuit around a dial.

Jimmy Soni and Rob Goodman, A Mind at Play