building-ai-agents

name	building-ai-agents
description	Guide for building AI agents from scratch with tool use, agentic loops, and LLM APIs. Includes nanocode as a complete working example showing how to build a minimal Claude Code alternative in 270 lines.

Building AI Agents

Complete guide for creating AI agents with tool-use capabilities from scratch.

When to Use

• Building new AI agents or coding assistants
• Implementing agentic loops with tool calling
• Learning agent architecture patterns
• Creating Claude API integrations

Core Architecture

Every AI agent needs four layers:

┌─────────────────────────────────────────┐
│  1. Tool Definition Layer              │  Define what tools exist
└─────────────────────────────────────────┘
                 ↓
┌─────────────────────────────────────────┐
│  2. Tool Implementation Layer         │  Implement tool logic
└─────────────────────────────────────────┘
                 ↓
┌─────────────────────────────────────────┐
│  3. Schema Generation Layer           │  Convert to API format
└─────────────────────────────────────────┘
                 ↓
┌─────────────────────────────────────────┐
│  4. Agentic Loop Layer               │  Execute until complete
└─────────────────────────────────────────┘

Quick Start

Minimal Template (30 lines)

import json, urllib.request

TOOLS = {
    "echo": (
        "Repeat input back",
        {"text": "string"},
        lambda args: args["text"],
    ),
}

def run_tool(name, args):
    return TOOLS[name][2](args)

def make_schema():
    return [{
        "name": name,
        "description": desc,
        "input_schema": {
            "type": "object",
            "properties": {
                k: {"type": v.rstrip("?").replace("number", "integer")}
                for k, v in params.items()
            },
            "required": [k for k, v in params.items() if not v.endswith("?")],
        }
    } for name, (desc, params, _) in TOOLS.items()]

while True:
    user_input = input("❯ ")
    response = urllib.request.urlopen(
        "https://api.anthropic.com/v1/messages",
        data=json.dumps({
            "model": "claude-3-5-sonnet-20241022",
            "max_tokens": 1024,
            "messages": [{"role": "user", "content": user_input}],
            "tools": make_schema()
        }).encode()
    )
    print(response.read())

30 lines. Working agent.

Key Principles

1. Dictionary-Based Parameters

API returns tool_use.input as JSON object → pass as dict:

args = block["input"]  # Already a dict!
result = tool(args)

2. Schema as Source of Truth

Generate from definition, never write twice:

TOOLS = {"tool": (desc, {"param": "type?"}, fn)}
# Schema auto-generates

3. Agentic Loop

Keep calling until no more tool uses:

while True:
    response = call_api(messages)
    tool_results = [execute(t) for t in response["tool_uses"]]

    if not tool_results:
        break  # Exit condition

    messages.append({"role": "user", "content": tool_results})

Complete Working Example

nanocode - Full Claude Code alternative in 270 lines:

• Location: examples/nanocode.py
• Tools: read, write, edit, glob, grep, bash
• Zero dependencies (Python stdlib only)
• Supports Anthropic API + OpenRouter

See: examples/nanocode.md

Common Patterns

Add New Tool

def my_tool(args):
    return process(args["input"])

TOOLS["my_tool"] = (
    "Description of what tool does",
    {"input": "string", "optional": "number?"},
    my_tool,
)

Optional Parameters

def read(args):
    offset = args.get("offset", 0)  # Default 0

Mark with ?: {"offset": "number?"}

Error Handling

def run_tool(name, args):
    try:
        return TOOLS[name][2](args)
    except Exception as err:
        return f"error: {err}"

Advanced Topics

Multi-step reasoning: references/multi-step-reasoning.md

Tool composition: references/tool-composition.md

State management: references/state-management.md

Troubleshooting: references/troubleshooting.md