AI Agent Examples: 12 Agents You Can Build This Weekend

👁 116 reads

An AI agent is a program where a language model chooses the next steps and tools to complete a task on its own. Below are 12 specific agents you can build in a weekend: each with a difficulty level, list of tools, and guidance on what to build it with (code or n8n/Make). We built and ran four of them on October 4, 2026, with results in separate articles. This is part of the AI Agents from Scratch learning path.

What Makes It an Agent vs Regular Automation

Automation executes steps in a fixed order: “email arrived, summarize it, save to spreadsheet”. An agent gets a goal and a set of tools, and the model chooses the order of steps. If data is incomplete, the agent can reach for an additional source, recalculate, or ask a human. We describe the difference and risks in more detail in AI Agent: What It Is and Where It Breaks.

In practice, most “agents” in companies are hybrids: a rigid workflow (e.g. in n8n) where one step is an agent with several tools. This is a good starting principle: use an agent only where steps truly depend on content.

12 AI Agents at a Glance

1. Multi-Currency Invoice Processing Agent (Difficulty 1, Built)

What it does: reads invoice files from a folder, recognizes amounts and currencies, fetches exchange rates from the central bank API, and calculates the total in local currency.

• Tools: search_files, read_page (central bank API), calculator.
• Test result: 3 steps, processed three invoices (including one for 120 EUR at rate 4.3745), cost $0.001. We tested it in Polish.
• Pitfall: models love to calculate in memory. The calculator description must say “always use this tool to calculate”.
• How to build: step by step in How to Build an AI Agent in Python.

2. Form Submission Triage Agent (Difficulty 1, Built)

What it does: contact form sends submission via webhook, model assigns category and urgency, writes draft response, and row goes to spreadsheet.

• Tools: webhook (FastAPI or Make module), model with JSON response, spreadsheet.
• Test result: 3 out of 3 submissions correctly categorized (invoices, quote request, complaint), 1.7-2.1 seconds and about $0.00016 per submission. We tested it in Polish.
• Pitfall: webhook without a secret is an open door for anyone who guesses the address.
• How to build: Webhook: What It Is and How to Connect AI to a Form. Similar scenario for email inbox: Email Automation with AI.

To be fair: this is more automation with one AI step than a full agent. It becomes an agent when you give the model tools like “check order status” and “find invoice” and let it decide what to use.

3. Company Document Chat Agent (Difficulty 2, Built)

What it does: answers questions based on your PDFs (policies, contracts, procedures) and cites which fragment the answer came from.

• Tools: embedding model (we used local Ollama), vector database (numpy or Chroma), response model.
• Test result: described in RAG Step by Step, tested on three documents in Polish.
• Pitfall: bad text chunking breaks search more than model choice does.

4. Business Days Calculator in Claude Code (Difficulty 1, Built)

What it does: custom MCP server gives Claude Code (or other client) tools: exchange rates, business day calculation with local holidays, note search.

• Test result: Claude calculated “14 business days from November 2, 2026” (November 20) on its own, calling the tool five times. We tested it in Polish.
• Pitfall: in Python SDK MCP 2.x, the FastMCP class was renamed to MCPServer.
• How to build: Your Own MCP Server in 15 Minutes.

5. Competitor Change Report (Difficulty 2, Pattern)

What it does: once a week visits competitor pricing pages, compares with previous version, and sends brief summary: what got more expensive, what’s new in the offer.

• Tools: page fetching, previous version storage (file or database), text comparison, email sending.
• Note: check the site’s terms of service and robots.txt. Don’t fetch pages more often than necessary.
• Where to start: read_page tool from agent #1 plus scheduler (Windows Task Scheduler, cron, or Schedule Trigger in n8n).

6. Meeting Notes to Task List (Difficulty 2, Pattern)

What it does: extracts tasks, people, and deadlines from transcription or notes, then creates them in Todoist, Trello, or Asana.

• Tools: transcription (local e.g. faster-whisper or your chat app’s function), model with JSON response, task list API.
• Pitfall: model adds tasks nobody assigned. Make it quote the sentence the task comes from.

7. Quote from Customer Inquiry (Difficulty 2, Pattern)

What it does: reads inquiry (“300 products, chatbot, November deadline”), selects items from price list, calculates amount, and assembles draft quote in template.

• Tools: CSV price list reading, calculator, template filling (e.g. text document or PDF).
• Pitfall: prices only from price list, never from model’s “memory”. Draft always reviewed by human.

8. Blog Post to Social Media (Difficulty 1, Pattern)

What it does: after blog post publication, fetches it from RSS, writes versions for LinkedIn, X, and Facebook in your style, and adds them to approval queue.

• Tools: RSS, model, post scheduler (Buffer, Metricool, or queue spreadsheet).
• Ready recipe: Social Media Automation with AI.

9. Code Fix Agent (Difficulty 2, Pattern)

What it does: receives bug report, reads repository, fixes code, runs tests, and prepares change for review.

• Tools: ready-made: Claude Code, Codex, Cursor. Custom: Claude Agent SDK with file and terminal tools.
• More: AI Agents for Code: Which to Choose and Codex vs Claude Code.

10. New Listings and Tenders Review (Difficulty 2, Pattern)

What it does: daily fetches new listings (jobs, tenders, contracts) from RSS feed or API, rejects non-matching ones, and sends 5 best with one-sentence justification.

• Tools: RSS or official source API, keyword filter, model for fit assessment, email or messenger.
• Note: use official data channels. Many sites prohibit automated fetching in their terms of service.

11. Research Assistant with Citations (Difficulty 3, Pattern)

What it does: receives question, plans searches on its own, reads several sources, saves notes, and delivers report with links for each claim.

• Tools: search API, page reading, notepad (file), optionally second agent as reviewer.
• Pitfall: cost grows fast because each read page goes into history. Summarize sources before adding to context. Framework choice: AI Agent Frameworks.

12. E-Commerce Customer Service (Difficulty 3, Pattern)

What it does: answers chat questions about shipping and returns (from FAQ and policies), checks order status, and when uncertain or customer is upset, hands off to human.

• Tools: RAG on FAQ and policies, store API (read-only), “transfer to consultant” tool.
• Legal note: customer must know they’re talking to AI. This follows from Article 50 of the AI Act, applicable from August 2, 2026. Details in AI Act Guide.

AI Agents in Polish: Does It Work

All four built agents received instructions and data in Polish and responded in Polish without problems: they understood local terms, exchange rates, business days, and correctly declined names. A few practical rules:

• write tool descriptions in your language, but function names without special characters (business_days, not business_days_holidays);
• data the model doesn’t know for sure (holidays, exchange rates, regulations) always provide via tool, don’t rely on its memory;
• when searching documents in languages other than English, choose a multilingual embedding model. Comparison of two models on Polish documents is in Embeddings and Vector Databases Explained;
• if data can’t leave the company, you can build some agents on local models, e.g. via Ollama.

How to Choose Your First Agent

1. Pick a task you do weekly that has a clear output (sum, category, list).
2. List what sources you use when doing it manually. These will be your tools.
3. Start with read-only tools. Add writing and sending only after the agent does well in logs for a week.
4. Calculate cost per run and multiply by monthly volume. Our simple agents cost fractions of a cent.

Common Questions

What’s the Simplest AI Agent Example?

An agent with a calculator and one data source, e.g. “calculate expenses from invoices at current exchange rate”. It has 2-3 tools, finishes in 2-3 steps, and it’s easy to verify if the result is correct.

Can You Build an AI Agent Without Coding?

Yes. n8n has an AI Agent node where you connect tools by clicking, and Gumloop walks you through it step by step. Examples of ready workflows: n8n in Practice.

How Much Does Running an AI Agent Cost?

Cost is model tokens plus any server. Our agents on a cheap model cost $0.00016 to $0.0013 per task. More expensive models and agents that read many pages cost multiples more.

How Does an AI Agent Differ from a Chatbot?

A chatbot responds to messages. An agent executes a task: it chooses tools on its own, checks results, and decides the next step. A chatbot with tools (like #12) is already an agent.

What Are Examples of AI Agents in Companies?

Most common: ticket and inbox handling, document preparation (quotes, reports), knowledge assistants on company documents, and code agents. Common denominator: repeatable task, access to company data, and human who approves the result.

Where to Find More Examples and Courses?

List of free courses and channels, verified October 4, 2026, is in the AI Agents from Scratch hub.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Verification Date

• Agents 1-4: our runs from October 4, 2026, code in the ai-agents-starter package.
• Patterns 5-12: based on Building effective agents (Anthropic) and AI Agents for Beginners course (Microsoft), verified October 4, 2026.

Kinetyra