1. Intro and Session Overview

• Presenter introduction "Hi everyone, I'm Nico and I work on the AI SDK at Vercel."
• Session goal "In this session, we'll explore how to build agents with the AI SDK."
• Structure
  • Fundamentals part: Introduction to the core building blocks of the AI SDK
  • Hands-on part: Building a Deep Research clone in Node.js

2. Project Setup and How to Run

• Clone the repo, install dependencies, copy environment variables "First, clone the repository, install the dependencies, then copy the environment variables."
• How to run
  • A single index.ts file
  • Run with pnpm rundev (alias pd)

3. Core Building Blocks: Text Generation Functions

3.1 generateText Function

• Basic usage

  • Call generateText inside an async main function
  • Model: OpenAI GPT-4.0 mini
  • Prompt: "hello world"
  • Print the result

    "Hello, how can I assist you today?"

• Message array support

  • You can pass a message array instead of a prompt
  • Messages are composed of role and content

• Advantage of a unified interface

  • "One of the key features of the AI SDK is its unified interface. You can switch between different models by changing just one line!"

3.2 Flexibility of Model Switching

• When web search is needed

  • GPT-4.0 mini only has training data up to 2024 and doesn't know about 2025 events

    "I'm sorry, but I don't have information about the AI Engineer Summit scheduled for 2025."

  • Switch to a Perplexity model
    • "Just swap the model."
    • Using Perplexity Sonar Pro

      "The AI Engineer Summit 2025 was held in New York from February 19 to 22."

    • Sources can be verified via the sources property

• Support for various providers

  • "You can find various providers in the SDK docs."
  • Example: Google Gemini Flash 1.5 + search grounding

4. Tools and Function Calling: Interacting with the Outside World

4.1 The Concept of Tools

• "Tools can seem complex at first glance, but they're simple at their core."

• How tools are provided

  • Pass a list of available tools to the model alongside the prompt
  • Each tool includes: name, description, and required data

• Tool calling flow

  • When the model determines a tool is needed, it generates a tool call
  • The developer parses that tool call, executes it, and processes the result

4.2 Simple Example: Adding Two Numbers

• Code structure

  • Pass a tools object to generateText
  • Tool name: addNumbers
  • Includes description, parameters, and execute function
  • Uses the tool utility for type safety

• Execution result

  "Tool result: addNumbers, args: 10, 5, result: 15"

• Only a tool call is generated, no text

  • "Now the model generates only a tool call, not text."

4.3 Synthesizing Tool Results into Text

• maxSteps option
  • "maxSteps causes the model to take the tool call result, feed it back into the conversation context, and repeat the next step."
  • "It repeats until text is generated or the maximum number of steps is reached."
  • Enables multi-step agent behavior

4.4 Demonstrating Agent Behavior with Multiple Tools

• Adding a weather information tool

  • getWeather tool: requires latitude, longitude, and city name
  • "If only the city name is in the prompt, the model can infer the remaining parameters."

• Compound prompt example

  "Get the weather for San Francisco and New York, then add the two values together."

• Execution result

  "The current temperature in San Francisco is 12.3°C and in New York is 15.2°C. Adding the two temperatures gives 27.5°C."

• Step structure

  • Step 1: Tool calls for weather in both cities
  • Step 2: Tool call to add the two values
  • Step 3: Final text generation

5. Structured Output

5.1 experimental output Option in generateText

• Defining a schema with Zod
  • "Zod is a TypeScript validation library that pairs perfectly with the AI SDK."
  • Example:

    output: object({
      sum: z.number(),
    });
    

• Execution result

  "experimental output: sum = 27.5"

5.2 generateObject Function

• Definition

  • "generateObject is a function specialized for structured output, and it's my favorite function."

• Example

  • Prompt: "Generate 10 definitions of an AI agent."
  • Schema: definitions: string[]
  • Result:

    "An AI agent is a software entity that performs tasks autonomously or semi-autonomously using artificial intelligence technology." (plus 9 more)

• Zod's describe feature

  • You can add detailed instructions per definition, such as "use as much jargon as possible and make it completely incomprehensible"
  • Result:

    "An autonomous entity leverages algorithmic heuristics to optimize decision-making processes in dynamic environments." (and so on)

6. Hands-on: Building a Deep Research Clone

6.1 What is Deep Research?

• "Deep Research is a service where you enter a topic, it browses the web, gathers materials, follows threads of thought, and ultimately produces a report."

• Core workflow

  1. Input query
  2. Generate sub-queries
  3. Web search for each query
  4. Analyze search results and generate follow-up questions
  5. Recursively repeat as needed
  6. Finally, synthesize all information into a report

• Depth and breadth concepts

  • "Depth: how many levels deep to go; breadth: how many queries to generate at each level"

6.2 Step-by-Step Implementation Details

1) Sub-query Generation Function

• generateSearchQueries
  • Input: query, number of queries to generate
  • Output: array of search-friendly queries
  • Example:

    "What does it take to become a D1 shot put athlete?"

    • "D1 shot put athlete requirements"
    • "D1 shot put athlete training methods"
    • "NCAA Division 1 shot put qualifications"

2) Web Search Function

• Using the Exa API
  • "Exa is a fast and affordable web search API."
  • Remove unnecessary information from results (favicons, etc.)
  • "Reducing the token count lowers costs and improves model efficiency."

3) Search Result Evaluation and Agent Loop

• searchAndProcess function
  • Uses two tools:
    1. Web search tool
    2. Search result evaluation tool (relevant/irrelevant)
  • Repeated using maxSteps
  • "If the evaluation result is irrelevant, it leaves feedback saying 'Please search again with a more specific query.'"
  • "This causes the model to keep trying to find better results."

4) Generating Learnings and Follow-up Questions

• generateLearnings function
  • Input: query, search results
  • Output:
    • learning (insight)
    • followUpQuestions (array of follow-up questions)
  • Example:

    "To become a D1 shot put athlete, you need four years of competitive experience in high school, placements at state/national competitions, and a throw of over 60 feet." "Follow-up questions: What are training methods for shot put athletes? What are the differences between D1, D2, and D3?"

5) Recursive Deep Research Function

• deepResearch function

  • Recursively uses followUpQuestions as new queries
  • Manages accumulated research state globally
  • Iterations limited by depth and breadth

• Preventing duplicate sources

  • "Sources already used are marked as irrelevant to prevent duplication."

6) Final Report Generation

• generateReport function

  • Generates a report based on accumulated research data
  • Uses the "o3 mini" model
  • Saves the result as a Markdown file

• Adding a system prompt

  • "You are a professional researcher. Today's date is XX. Use Markdown format. Clearly label any speculation or predictions."
  • "To prevent the model from relying on reasoning alone, clearly specify the structure and format you want."

• Execution result

  "Below is a comprehensive report on the requirements, skills, training, common beginner mistakes, and more for becoming a D1 shot put athlete."

  • "Top-level competitiveness, four years of athletic experience, throws of 55 feet or more, differences by gender, and more..."

7. Closing and Resources

• "That's how you can build a Deep Research agent in just 218 lines of code!"
• "If you have questions, DM me on X (Twitter) @nikoalbanese10."
• "Also check out the docs and cookbook guides at SDK.purcell.ai."
• "Thanks to Swix for suggesting this session, and see you next time!"

Key Keyword Summary

• AI SDK
• generateText / generateObject
• Tools / function calling
• Multi-step agents
• Structured output
• Zod schema
• Deep Research workflow
• Recursive agents
• Web search API (Exa)
• Markdown report generation
• System prompt design

"I hope this session inspires you to go build your own amazing agents and research systems with the AI SDK!" 😊