The Minimal Agent Loop — Build Your First Agent

Your First Agent in 50 Lines

Forget frameworks. Build one from scratch to truly understand agents.

The Architecture

┌────────────────────────────────────┐
│       THE MINIMAL AGENT LOOP       │
├────────────────────────────────────┤
│  1. Send messages + tools to LLM   │
│  2. Get response                   │
│  3. If response has tool_use:      │
│     a. Execute the tool             │
│     b. Add result to messages       │
│     c. GOTO step 1                  │
│  4. If response has text:           │
│     a. Return the text (DONE)       │
└────────────────────────────────────┘

Complete Implementation (TypeScript)

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();
const tools = [{
  name: "get_weather",
  description: "Get current weather for a city",
  input_schema: {
    type: "object",
    properties: {
      city: { type: "string", description: "City name" }
    },
    required: ["city"]
  }
}];

async function runAgent(userMessage: string) {
  const messages = [{ role: "user", content: userMessage }];
  
  while (true) {
    const response = await client.messages.create({
      model: "claude-sonnet-5-5",
      max_tokens: 1024,
      tools,
      messages
    });
    
    if (response.stop_reason === "tool_use") {
      const toolBlock = response.content.find(b => b.type === "tool_use");
      const result = executeWeather(toolBlock.input.city);
      messages.push({ role: "assistant", content: response.content });
      messages.push({
        role: "user",
        content: [{ type: "tool_result", tool_use_id: toolBlock.id,
                     content: JSON.stringify(result) }]
      });
    } else {
      return response.content[0].text; // Done!
    }
  }
}

Production Shortcut: Anthropic Claude Agent SDK

Once you understand the raw while loop, Anthropic's official Claude Agent SDK (pip install claude-agent-sdk in Python and npm install @anthropic-ai/claude-agent-sdk in TypeScript - renamed from claude-code-sdk / @anthropic-ai/claude-code) packages this exact loop with built-in file/bash/MCP tools, context compaction, and permission gates:

from claude_agent_sdk import query, ClaudeAgentOptions

async for message in query(
    prompt="Inspect src/ and summarize potential race conditions.",
    options=ClaudeAgentOptions(
        allowed_tools=["Read", "Grep", "Glob"],
        permission_mode="default",
        max_turns=6,
    ),
):
    print(message)