The Agent Loop

[ s01 ] > s02 > s03 > s04 > s05 > s06 > s07 > s08 > s09 > s10 > s11 > s12 > s13 > s14 > s15 > s16 > s17 > s18 > s19

What You'll Learn

• How the core agent loop works: send messages, run tools, feed results back
• Why the "write-back" step is the single most important idea in agent design
• How to build a working agent in under 30 lines of Python

Imagine you have a brilliant assistant who can reason about code, plan solutions, and write great answers -- but cannot touch anything. Every time it suggests running a command, you have to copy it, run it yourself, paste the output back, and wait for the next suggestion. You are the loop. This chapter removes you from that loop.

The Problem

Without a loop, every tool call requires a human in the middle. The model says "run this test." You run it. You paste the output. The model says "now fix line 12." You fix it. You tell the model what happened. This manual back-and-forth might work for a single question, but it falls apart completely when a task requires 10, 20, or 50 tool calls in a row.

The solution is simple: let the code do the looping.

The Solution

Here's the entire system in one picture:

+--------+      +-------+      +---------+
|  User  | ---> |  LLM  | ---> |  Tool   |
| prompt |      |       |      | execute |
+--------+      +---+---+      +----+----+
                    ^                |
                    |   tool_result  |
                    +----------------+
                    (loop until the model stops calling tools)

The model talks, the harness (the code wrapping the model) executes tools, and the results go right back into the conversation. The loop keeps spinning until the model decides it's done.

How It Works

Step 1. The user's prompt becomes the first message.

// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }
  anthropicClient = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_AUTH_TOKEN || "missing-api-key",
    baseURL: process.env.ANTHROPIC_BASE_URL || void 0
  });
  return anthropicClient;
}
function createLoopContext() {
  return { workdir: WORKDIR, messages: [], meta: {} };
}
function safePath(relativePath) {
  const resolved = path.resolve(WORKDIR, relativePath);

Step 2. Send the conversation to the model, along with tool definitions.

// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }
  anthropicClient = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_AUTH_TOKEN || "missing-api-key",
    baseURL: process.env.ANTHROPIC_BASE_URL || void 0
  });
  return anthropicClient;
}
function createLoopContext() {
  return { workdir: WORKDIR, messages: [], meta: {} };
}
function safePath(relativePath) {
  const resolved = path.resolve(WORKDIR, relativePath);

Step 3. Add the model's response to the conversation. Then check: did it call a tool, or is it done?

// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }
  anthropicClient = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_AUTH_TOKEN || "missing-api-key",
    baseURL: process.env.ANTHROPIC_BASE_URL || void 0
  });
  return anthropicClient;
}
function createLoopContext() {
  return { workdir: WORKDIR, messages: [], meta: {} };
}
function safePath(relativePath) {
  const resolved = path.resolve(WORKDIR, relativePath);

Step 4. Execute each tool call, collect the results, and put them back into the conversation as a new message. Then loop back to Step 2.

// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }
  anthropicClient = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_AUTH_TOKEN || "missing-api-key",
    baseURL: process.env.ANTHROPIC_BASE_URL || void 0
  });
  return anthropicClient;
}
function createLoopContext() {
  return { workdir: WORKDIR, messages: [], meta: {} };
}
function safePath(relativePath) {
  const resolved = path.resolve(WORKDIR, relativePath);

Put it all together, and the entire agent fits in one function:

// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }
  anthropicClient = new Anthropic({
    apiKey: process.env.ANTHROPIC_API_KEY || process.env.ANTHROPIC_AUTH_TOKEN || "missing-api-key",
    baseURL: process.env.ANTHROPIC_BASE_URL || void 0
  });
  return anthropicClient;
}
function createLoopContext() {
  return { workdir: WORKDIR, messages: [], meta: {} };
}
function safePath(relativePath) {
  const resolved = path.resolve(WORKDIR, relativePath);

That's the entire agent in under 30 lines. Everything else in this course layers on top of this loop -- without changing its core shape.

A note about real systems: Production agents typically use streaming responses, where the model's output arrives token by token instead of all at once. That changes the user experience (you see text appearing in real time), but the fundamental loop -- send, execute, write back -- stays exactly the same. We skip streaming here to keep the core idea crystal clear.

What Changed

Try It

npm run s01

1. Ask the agent to run pwd
2. Ask it to run ls -la
3. Ask it to summarize the current workspace in one sentence
4. Ask it to create notes/hello.ts and print the file content

What You've Mastered

At this point, you can:

• Build a working agent loop from scratch
• Explain why tool results must flow back into the conversation (the "write-back")
• Redraw the loop from memory: messages -> model -> tool execution -> write-back -> next turn

What's Next

Right now, the agent can only run bash commands. That means every file read uses cat, every edit uses sed, and there's no safety boundary at all. In the next chapter, you'll add dedicated tools with a clean routing system -- and the loop itself won't need to change at all.

Key Takeaway

An agent is just a loop: send messages to the model, execute the tools it asks for, feed the results back, and repeat until it's done.

---

TypeScript Companion (This Repo)

• Core addition: LoopState + tool_result feedback
• Key insight: An agent is just a loop: send messages, execute tools, feed results back, repeat.
• File: s01_agent_loop.ts
• LOC: 393

npm run s01

// @ts-nocheck
// AUTO-GENERATED by scripts/generate-self-contained-agents.mjs
// Source chapter: agents_self_contained/s01_agent_loop.ts
// Tree-shaken bundle: chapter wiring + only used runtime code.
// agents_self_contained/_runtime.ts
import Anthropic from "@anthropic-ai/sdk";
import dotenv from "dotenv";
import { execSync, spawn, spawnSync } from "node:child_process";
import fs from "node:fs";
import path from "node:path";
import process from "node:process";
import readline from "node:readline/promises";
import { stdin as input, stdout as output } from "node:process";
dotenv.config({ override: true });
var WORKDIR = process.cwd();
var DEFAULT_MODEL = "claude-3-5-sonnet-latest";
var anthropicClient = null;
function getModelId() {
  return process.env.MODEL_ID || DEFAULT_MODEL;
}
function getAnthropicClient() {
  if (anthropicClient) {
    return anthropicClient;
  }