How to Build an AI Agent in 2026: A Practical Step-by-Step Guide

Most "how to build an AI agent" tutorials start with installing LangChain and 47 dependencies. Then you spend 3 hours debugging import errors before writing a single line of useful code.

This guide is different. We'll build a real, working AI agent from scratch using nothing but Python and an API key. No frameworks. No vector databases. No Kubernetes. Just a script that runs, makes decisions, and does useful work.

By the end, you'll have a fully autonomous agent that can browse the web, analyze content, and take actions — the same pattern behind tools like Devin, Cursor, and Claude Code.

What Is an AI Agent (Really)?

An AI agent isn't a chatbot. A chatbot waits for your input and responds. An agent acts on its own.

The difference is a loop. Here's the core pattern every AI agent follows:

while not done:
    observation = perceive(environment)
    thought = think(observation, goal)
    action = decide(thought)
    result = execute(action)
    done = evaluate(result, goal)

That's it. Every AI agent — from a $50M startup's product to a weekend project — follows this loop. The LLM is the "think" step. Everything else is just code you already know how to write.

1 Set Up Your Environment

You need Python 3.10+ and an API key from any major LLM provider. We'll use examples with multiple providers so you can pick what fits your budget:

Provider	Model	Cost per 1M tokens	Best for
Anthropic	Claude Sonnet 4.6	$3 / $15	Complex reasoning
OpenAI	GPT-5.4	$2.50 / $10	General purpose
DeepSeek	V3.2	$0.27 / $1.10	Cost-sensitive apps
Google	Gemini 3.1 Flash	$0.25 / $1.25	High throughput

# Install the only dependency you need
pip install requests

# That's literally it. No langchain, no llamaindex, no crew.

2 Build the Core Agent Loop

Here's a minimal but complete AI agent in ~50 lines of Python. This agent can use tools, reason about results, and decide when it's done:

import json
import requests

API_KEY = "your-api-key-here"
API_URL = "https://api.anthropic.com/v1/messages"

TOOLS = {
    "search_web": lambda q: requests.get(
        f"https://api.duckduckgo.com/?q={q}&format=json"
    ).json().get("AbstractText", "No results"),

    "read_url": lambda url: requests.get(url, timeout=10).text[:3000],

    "save_file": lambda content: open("output.txt", "w").write(content) or "Saved!",
}

SYSTEM = """You are an autonomous agent. You have these tools:
- search_web(query): Search the web
- read_url(url): Read a webpage
- save_file(content): Save content to a file

Respond with JSON: {"thought": "...", "tool": "tool_name", "args": "...", "done": false}
When finished: {"thought": "...", "done": true, "answer": "..."}"""

def call_llm(messages):
    resp = requests.post(API_URL, json={
        "model": "claude-sonnet-4-6-20250514",
        "max_tokens": 1024,
        "system": SYSTEM,
        "messages": messages,
    }, headers={
        "x-api-key": API_KEY,
        "anthropic-version": "2023-06-01",
        "content-type": "application/json",
    })
    return resp.json()["content"][0]["text"]

def run_agent(goal, max_steps=10):
    messages = [{"role": "user", "content": f"Goal: {goal}"}]

    for step in range(max_steps):
        response = call_llm(messages)
        print(f"\n--- Step {step + 1} ---")
        print(response)

        data = json.loads(response)

        if data.get("done"):
            return data.get("answer", "Done")

        # Execute the tool
        tool_fn = TOOLS.get(data["tool"])
        result = tool_fn(data["args"]) if tool_fn else "Unknown tool"

        messages.append({"role": "assistant", "content": response})
        messages.append({"role": "user", "content": f"Tool result: {result}"})

    return "Max steps reached"

# Run it
answer = run_agent("Find the latest news about AI agents and summarize the top 3 stories")
print(f"\nFinal answer: {answer}")

That's a complete agent. ~50 lines. It observes (tool results), thinks (LLM), decides (JSON output), and acts (tool execution). No framework needed.

3 Add Memory

Agents without memory forget everything between steps. There are two types of memory you need:

Short-term memory (conversation context)

This is already built into the code above — the messages list carries the full conversation. But it grows fast and eats tokens. Add a simple summary mechanism:

def compress_memory(messages, keep_last=4):
    """Summarize old messages to save tokens."""
    if len(messages) <= keep_last:
        return messages

    old = messages[:-keep_last]
    summary = call_llm([{
        "role": "user",
        "content": f"Summarize this conversation concisely:\n{json.dumps(old)}"
    }])

    return [{"role": "user", "content": f"Previous context: {summary}"}] + messages[-keep_last:]

Long-term memory (persistent storage)

For agents that run across sessions, save key facts to a JSON file:

import json
from pathlib import Path

MEMORY_FILE = Path("agent_memory.json")

def remember(key, value):
    memory = json.loads(MEMORY_FILE.read_text()) if MEMORY_FILE.exists() else {}
    memory[key] = value
    MEMORY_FILE.write_text(json.dumps(memory, indent=2))

def recall(key):
    memory = json.loads(MEMORY_FILE.read_text()) if MEMORY_FILE.exists() else {}
    return memory.get(key)

4 Make It Reliable

Real agents need to handle failures gracefully. The #1 cause of agent failures is the LLM returning malformed output. Here's the fix:

import re

def parse_response(text):
    """Extract JSON from LLM response, even if surrounded by text."""
    # Try direct parse first
    try:
        return json.loads(text)
    except json.JSONDecodeError:
        pass

    # Find JSON block in response
    match = re.search(r'\{[\s\S]*\}', text)
    if match:
        try:
            return json.loads(match.group())
        except json.JSONDecodeError:
            pass

    # Fallback: ask the LLM to fix it
    return {"thought": "Failed to parse response", "done": True, "answer": text}

def safe_execute(tool_fn, args, timeout=30):
    """Execute a tool with error handling."""
    try:
        return str(tool_fn(args))
    except Exception as e:
        return f"Error: {type(e).__name__}: {e}"

5 Add Useful Tools

The power of an agent comes from its tools. Here are the most useful ones you can add in a few lines each:

import subprocess
import sqlite3

TOOLS = {
    # Web browsing
    "search_web": lambda q: requests.get(
        f"https://api.duckduckgo.com/?q={q}&format=json"
    ).json().get("AbstractText", "No results"),

    "read_url": lambda url: requests.get(url, timeout=10).text[:5000],

    # File operations
    "read_file": lambda path: Path(path).read_text()[:5000],
    "write_file": lambda args: Path(args["path"]).write_text(args["content"]),
    "list_files": lambda dir: str(list(Path(dir).glob("*"))),

    # Code execution (sandboxed)
    "run_python": lambda code: subprocess.run(
        ["python3", "-c", code],
        capture_output=True, text=True, timeout=30
    ).stdout[:2000],

    # Database
    "query_db": lambda sql: str(
        sqlite3.connect("data.db").execute(sql).fetchall()[:20]
    ),

    # Send notification
    "notify": lambda msg: requests.post(
        f"https://api.telegram.org/bot{BOT_TOKEN}/sendMessage",
        json={"chat_id": CHAT_ID, "text": msg}
    ).json(),
}

6 Run It on a Schedule

An agent that only runs when you manually start it isn't really autonomous. Here's how to make it run on its own:

Option A: Cron job (simplest)

# Run every day at 8am
# Edit with: crontab -e
0 8 * * * cd /path/to/agent && python3 agent.py >> agent.log 2>&1

Option B: Continuous loop with sleep

import time

def main_loop():
    while True:
        try:
            result = run_agent("Check for new events and take action")
            print(f"[{time.strftime('%H:%M')}] {result}")
        except Exception as e:
            print(f"Error: {e}")
        time.sleep(3600)  # Run every hour

if __name__ == "__main__":
    main_loop()

Option C: Event-driven (webhook)

from flask import Flask, request

app = Flask(__name__)

@app.route("/webhook", methods=["POST"])
def webhook():
    data = request.json
    result = run_agent(f"Handle this event: {json.dumps(data)}")
    return {"status": "ok", "result": result}

# Run with: python3 -m flask run --port 5000

7 Real-World Example: News Monitoring Agent

Let's put it all together. Here's a complete agent that monitors tech news and sends you a daily digest — the same pattern we use to run AI Agents Weekly:

#!/usr/bin/env python3
"""Autonomous news monitoring agent."""
import json, requests, time
from datetime import datetime

API_KEY = "your-key"
SOURCES = [
    "https://hn.algolia.com/api/v1/search?query=AI+agent&tags=story",
    "https://hn.algolia.com/api/v1/search?query=LLM+autonomous&tags=story",
]

def fetch_news():
    articles = []
    for url in SOURCES:
        try:
            data = requests.get(url, timeout=10).json()
            for hit in data.get("hits", [])[:5]:
                articles.append({
                    "title": hit["title"],
                    "url": hit.get("url", ""),
                    "points": hit.get("points", 0),
                })
        except Exception:
            continue
    return sorted(articles, key=lambda x: x["points"], reverse=True)[:10]

def summarize(articles):
    text = "\n".join(f"- {a['title']} ({a['points']} pts)" for a in articles)
    resp = requests.post("https://api.anthropic.com/v1/messages", json={
        "model": "claude-sonnet-4-6-20250514",
        "max_tokens": 500,
        "messages": [{"role": "user", "content":
            f"Summarize these top AI stories in 3 bullet points:\n{text}"}],
    }, headers={
        "x-api-key": API_KEY,
        "anthropic-version": "2023-06-01",
        "content-type": "application/json",
    })
    return resp.json()["content"][0]["text"]

def send_digest(summary, articles):
    msg = f"🤖 AI Agent Daily Digest — {datetime.now().strftime('%b %d')}\n\n"
    msg += summary + "\n\n📰 Top stories:\n"
    msg += "\n".join(f"• {a['title']}\n  {a['url']}" for a in articles[:5])
    # Send via Telegram, email, Slack, etc.
    print(msg)

if __name__ == "__main__":
    articles = fetch_news()
    summary = summarize(articles)
    send_digest(summary, articles)

Framework vs. No Framework: When to Use What

We just built agents without any framework. So when should you use one?

Use a framework when...	Skip the framework when...
Multiple agents need to collaborate	You have a single agent
You need complex state machines	A simple loop is enough
You want built-in observability	Print statements work fine
Your team needs standardized patterns	You're a solo builder
You need production-grade retries/fallbacks	try/except covers your needs

If you do want a framework, check out our comparison of the top 7 AI agent frameworks in 2026.

Common Mistakes to Avoid

• Too many tools: Start with 3-5 tools. More tools = more confusion for the LLM. Add tools only when the agent fails because it lacks a capability.
• No output validation: Always parse and validate the LLM's JSON output. It will break eventually.
• Infinite loops: Set a max_steps limit. Agents that run forever cost real money.
• No logging: Log every step. When your agent does something unexpected at 3am, you need to know why.
• Over-engineering early: Get a working prototype first. Add vector search, RAG, and multi-agent orchestration only when you actually need them.

What's Next

You now have a working AI agent. Here's what to explore next:

1. Add RAG — connect your agent to a document database for domain-specific knowledge
2. Multi-agent systems — have specialized agents collaborate (one researches, one writes, one reviews)
3. Human-in-the-loop — add approval steps before high-stakes actions
4. Evaluation — build test cases to measure your agent's accuracy over time

The AI agent space moves fast. We cover the latest tools, frameworks, and deployment patterns in AI Agents Weekly — free, 3x/week, no fluff.

Related Articles

• How to Run Autonomous AI Agents with Claude Code
• What Is MCP (Model Context Protocol)?
• Top 7 AI Agent Frameworks in 2026
• What Are AI Agents? The Complete Guide