Building an AI Agent from Scratch Without Frameworks

Summary

Dennis Stanoev advocates learning agent fundamentals before adopting frameworks. The core insight: an agent is just a wrapper around an LLM that allows the model to make decisions and take actions. By building a customer support agent from scratch with TypeScript and the Gemini API, he demonstrates that the architecture is surprisingly simple.

Key Concepts

The Agent Loop

Every agent follows a five-step pattern:

1. Define available tools with metadata (parameters, descriptions)
2. Send user prompt and tool definitions to the LLM
3. Model analyzes request and returns structured tool-use responses
4. Execute intercepted tool calls and capture results
5. Return results to model for next iteration or final response

Static vs Dynamic Content

FAQ information belongs in system prompts rather than tool calls. Static content embedded in prompts reduces latency compared to fetching through tools.

Function Declarations

Tools are defined through function declarations—metadata that tells the LLM what the tool does and what parameters it accepts. The LLM invokes tools; handlers execute actual operations.

Visual Model

flowchart LR
    Input[User Message] --> LLM[Gemini API]
    LLM --> Check{Tool Call?}
    Check -->|Yes| Exec[Execute Tool]
    Exec --> Result[Return Results]
    Result --> LLM
    Check -->|No| Done[Final Response]

Code Snippets

Tool Definition

How to define a tool for the Gemini API.

const validateOrderDeclaration: FunctionDeclaration = {
  name: "validateOrder",
  description: "Validates an order by its ID and returns order details",
  parameters: {
    type: SchemaType.OBJECT,
    properties: {
      orderId: {
        type: SchemaType.STRING,
        description: "The unique identifier of the order",
      },
    },
    required: ["orderId"],
  },
};

Agent Loop

Core loop that powers the agent.

while (true) {
  const result = await chat.sendMessage(userMessage);
  const response = result.response;

  const functionCalls = response.functionCalls();
  if (!functionCalls || functionCalls.length === 0) {
    // No tool calls - we have our final response
    return response.text();
  }

  // Execute each tool call and collect results
  for (const call of functionCalls) {
    const toolResult = await executeTool(call.name, call.args);
    // Send result back to continue the conversation
  }
}

Connections

• how-to-build-an-agent - Thorsten Ball arrives at the same "loop until done" pattern using Go and the Anthropic API, reinforcing that this architecture is universal
• building-effective-agents - Anthropic's official guide provides complementary workflow patterns and production considerations for when the basic loop needs orchestration