Master AI Agents
on n8n from zero
and start earning

n8n — documentation & templates

YouTube — best channels

Courses & guides — agents

Prompt engineering

Communities & newsletters

---

Don't touch any tools in week one. Read, watch, draw diagrams. The goal is a solid mental model: what an agent is, how it makes decisions, what tools and memory mean in practice.

W1 · Days 1–7 How AI agents work — the mental model Mar 8–14 0%

Read and think. No interfaces, no clicking. A clear architecture diagram in your head is the only deliverable this week.

• Read "Building Effective Agents" by Anthropic — the definitive guide to agent architecture (30 min)
• Understand the core difference: ChatGPT = one request → one response. Agent = a loop: think → choose tool → get result → think again
• Study the 5 agentic patterns from the Anthropic article: Prompt Chaining, Routing, Parallelization, Orchestrator-Workers, Evaluator-Optimizer — sketch each one on paper
• Watch "What are AI Agents?" — Andrew Ng on YouTube (20 min). Write down 5 key insights in your own words
• Complete lessons 1–3 of Microsoft AI Agents for Beginners: Introduction, Frameworks, Design Patterns
• Draw an agent map: LLM core → Tools → Short-term memory (context) → Long-term memory (database) → Output
• Read about Function Calling in OpenAI Docs — this is the mechanism agents use to invoke tools

Weekly challenge

Explain in 3 minutes — to a friend or to yourself — what an AI agent is, how it differs from ChatGPT, and why tools matter. No jargon. If you can do it, you've got it.

W2 · Days 8–14 Setting up n8n — first workflow running Mar 15–21 0%

Install n8n, connect the OpenAI API, build your first working workflow. By the end of the week — a Telegram bot that answers questions.

• Sign up at platform.openai.com → API Keys → create a key → add $5–10 to your balance
• Sign up at n8n.io — cloud version is free for 14 days. Complete the built-in tutorial (20 min)
• Build your first workflow: Webhook → OpenAI Chat → Respond. Send a POST request via Hoppscotch or curl — the agent replied
• Watch one Leon Van Zyl video — "Build an AI Agent with n8n" — and follow along step by step
• Connect Telegram: create a bot via @BotFather → add Telegram Trigger node in n8n → the bot now responds via OpenAI
• Read the n8n AI Documentation — understand which AI nodes exist and what each one does
• Browse 20+ templates in n8n AI Templates — write down 3 agent ideas you want to build

Weekly challenge

A working Telegram bot: you type a question → the bot answers via OpenAI. Zero lines of code. Show it to a friend and get "how does that work?"

W3 · Days 15–20 Agent with tools — the full loop Mar 22–27 0%

An agent without tools is just a chat. Tools give the agent the ability to act: search the web, read files, send messages, run calculations.

• In n8n, use the "AI Agent" node (not just OpenAI Chat) — this is a full ReAct agent that can invoke tools in a loop
• Add tools to the agent: Wikipedia (fact lookup) + Calculator + HTTP Request (any external API)
• Connect SerpAPI (free plan) — the agent can now search Google for current data. Ask about something happening today
• Add memory: the Window Buffer Memory node — the agent now remembers previous messages in the conversation
• Build the first complete workflow: Telegram → AI Agent (search + memory) → formats the answer → sends it back to Telegram
• Set up error handling: add an Error Trigger → Telegram notification if the workflow fails

Weekly challenge

Give the agent a topic → it searches the web → writes a structured report → sends it to Telegram. Zero involvement from you once triggered.

---

Prompts are the new code. The difference between an agent that works and one that "sort of works" is the System Prompt. One well-crafted prompt can replace 200 lines of logic.

W4 · Days 21–28 Anatomy of a prompt — from bad to great Mar 28 – Apr 4 0%

Learn to write System Prompts that make agents behave predictably. Structure, examples, constraints, output format.

• Complete chapters 1–3 of Learn Prompting: Basics, Advanced Techniques, Chain of Thought (2–3 hrs, entirely free)
• Read the OpenAI Prompt Engineering Guide — write down 10 specific rules you'll apply immediately
• Write a System Prompt using this structure: Role → Context → Goals → Constraints → Output format → Examples. Minimum 300 words for a test agent
• Experiment in OpenAI Playground: change temperature, system prompt, add few-shot examples — document how each change affects output
• Apply Structured Output in n8n: agent returns JSON → Code node parses it → data flows to the next step automatically
• Start a "prompt library" in Notion — categories: assistant, analysis, content generation, classification, structured output. Add every good prompt you write

Weekly challenge

Take one agent. Write three versions of the System Prompt — bad, average, excellent. Compare responses to the same questions. Document the difference.

W5 · Days 29–35 Agent-level prompts — advanced techniques Apr 5–11 0%

Agent prompts are more complex than chat prompts — the agent makes decisions, invokes tools, evaluates results. The prompt must govern the entire loop.

• Read about ReAct prompting on DAIR.AI Prompting Guide — write a ReAct prompt for an agent with three tools
• Build a Routing prompt in n8n: a dispatcher agent reads incoming requests and routes them via a Switch node to the correct sub-workflow
• Write an Evaluator prompt: a separate agent checks another agent's answer against criteria (accuracy, completeness, format). The self-checking pattern
• Study Prompt Injection — how input data can override agent behaviour. Add a defensive layer in the System Prompt
• Complete Prompt Engineering for Developers — Isa Fulford + Andrew Ng, DeepLearning.AI, free, 2 hrs

# System Prompt template for an agent You are [ROLE]. You operate in this context: [CONTEXT]. Goals: 1. [PRIMARY GOAL] 2. [SECONDARY GOAL] Rules: - Always [REQUIRED ACTION] - Never [PROHIBITION] - If [CONDITION] → [RESPONSE] Output format: JSON / Markdown / Plain text Examples: Input: [EXAMPLE 1] Output: [EXPECTED RESPONSE 1]

Weekly challenge

A library of 10 working prompts for different agent types. Each one tested on at least 5 questions and refined iteratively.

---

W6 · Days 36–44 RAG — agent that knows your data Apr 12–20 0%

RAG = the agent answers based on your documents, not just its training data. The core technology behind enterprise AI agents. Done visually in n8n.

• Understand the RAG pipeline: Document → Chunks → Embeddings → Vector DB → Retrieval → Answer. Draw the diagram
• In n8n: build a document ingestion pipeline → Default Data Loader → Recursive Character Text Splitter → Embeddings OpenAI → Qdrant Vector Store
• Build an agent that answers questions from loaded docs: AI Agent + Vector Store Tool → ask it questions about the uploaded text
• Load 10+ real documents (manuals, FAQs, articles) and verify the agent answers more accurately than vanilla GPT-4
• Add persistent memory via Postgres or Supabase — the agent remembers data between sessions and across users

Weekly challenge

Topic-expert agent: load 20+ documents on any subject you care about. The agent should outperform generic ChatGPT because it has your specific data.

W7 · Days 45–52 Automating the routine — three working agents Apr 21–28 0%

An agent that saves 2–4 hours a day is one people pay for. Build automations with a measurable result.

• Content agent: daily Cron trigger → reads RSS feeds → picks top 5 relevant stories → writes a post → sends for approval in Telegram
• Email agent: Gmail Trigger → classifies emails → handles urgent ones → drafts replies → waits for your confirmation before sending
• Monitoring agent: HTTP Request every hour → scrapes competitor prices / job listings / news → sends a Telegram alert on change
• Add Human-in-the-loop: agent pauses, sends "Approve?" → waits for a button press in Telegram via Wait node → continues
• Measure: how many minutes does each agent save per day? Record the numbers — you'll need them when pricing your services

Weekly challenge

Three working agents that together save you 2+ hours a day. Each deployed on Railway.app and running 24/7.

W8 · Days 53–60 Portfolio — three agents ready to show Apr 29 – May 6 0%

"Show me" convinces 10× more than "let me tell you". Three live agents people can actually try are your best argument to clients.

• Deploy n8n self-hosted on Railway.app (free tier) — agents run 24/7 without your computer
• Record a Loom demo for each agent: "Problem → What the agent does → Result." No technical details — only value. 3–5 minutes each
• Build a portfolio page in Notion: three case studies with metrics ("saves 2 hrs/day", "handles 50 requests/day")
• Give access to five entrepreneurs you know. Record their reactions — what they liked, what was confusing, what they'd change

Weekly challenge

Three live agents, three Loom videos, one portfolio page. Show ten people and get three "how much does this cost?"

---

W9 · Days 61–68 Multi-agent — orchestrator and subagents May 7–14 0%

One agent = one specialist. Multiple agents = an entire department. The orchestrator receives a task, delegates to subagents, collects the result.

• Study the Orchestrator-Workers pattern from the Anthropic article. Sketch a system: 1 orchestrator + 3–4 specialised subagents
• In n8n: "AI editorial team" — orchestrator splits the task → Research agent (web search) → Writer agent (text) → Editor agent (review) → final output
• Use the Execute Workflow node to connect agents — the primary tool for multi-agent architecture in n8n
• Add parallel processing: multiple subagents run simultaneously via SplitInBatches — results are merged downstream
• Build an "AI marketing department": topic → competitor research → strategy → 5 posts → publishing schedule. Four agents, one trigger

Weekly challenge

A working system of four agents. One input → orchestrator → four subagents → finished output. Not a single line of code except maybe a simple JSON parse.

W10 · Days 69–74 Voice & Vision May 15–20 0%

Voice agents are the fastest-growing segment of 2025–2026. Vision agents analyse images. n8n becomes the processing hub for both.

• Sign up at Vapi.ai (free tier) — create a voice agent in 20 minutes. Call it. Connect n8n as a webhook for business logic
• Build the full voice loop: call → Vapi transcribes → webhook to n8n → AI Agent processes → Vapi reads the response aloud
• Vision in n8n: Telegram sends a photo → OpenAI Vision node analyses it → agent describes, classifies, or extracts data from the image
• Practical Vision case: agent reads a photo of a receipt/invoice → extracts data (total, date, line items) → logs it to Airtable automatically

Weekly challenge

A voice agent that answers calls, captures enquiries and saves them to a spreadsheet. Show three people — get "how much does this cost?"

W11 · Days 75–80 Reliability — from hobby to product May 21–26 0%

An agent that works 95% of the time is a hobby. One that works 99.9% is a product. The difference: error handling, monitoring, and cost optimisation.

• Set up monitoring: Error Trigger → Telegram alert with error details. Add this to every production workflow
• Add retry logic: if the API returns an error — wait 30 seconds and try again. Wait node + attempt counter
• Optimise costs: gpt-4o-mini for simple steps (10× cheaper), gpt-4o only for complex reasoning. Calculate monthly cost per agent
• Build a simple dashboard in Airtable or Notion: requests processed, cost per day, % successful responses. Foundation for pricing

Weekly challenge

Take your best agent. Bring it to production quality: error handling, monitoring, cost optimisation. Write down: running cost per month = $X. That's your pricing floor.

---

W12 · Days 81–87 First clients May 27 – Jun 2 0%

The fastest path to first money: find a client with a problem → solve it with an agent → get paid → use it as a case study. Repeat.

• Create profiles on Upwork and Fiverr — service: "AI automation / n8n workflow building". First orders start here
• Run 5 free "AI audits": listen to what a business owner does manually → show how an agent can automate it → propose a solution. This is the best sales technique
• Write an ROI calculator: "You spend X hours on task Y. The agent does it in Z minutes. Savings: $N/month." This kills pricing objections
• Take the first project — even for modest pay. Do it brilliantly. Get a written testimonial and permission to use it as a case study. The case study is worth more than the fee.
• Create a simple landing page on Carrd.co (free) or Notion: portfolio + services + pricing + contact form

Weekly challenge

Five real conversations with potential clients. Don't sell — listen. What hurts? What's repeated every day? What wastes most time? That's your next agent.

W13 · Days 88–94 Content & launch Jun 3–9 0%

Content is the cheapest B2B channel. One honest case study keeps working for months. Build a funnel that sells while you sleep.

• Write a case study: "How I went from zero to building AI agents in 100 days" — publish on dev.to or Medium. Honesty + real numbers = traffic
• Start a LinkedIn content series: three posts per week — tip, case study, tutorial. Target: 500 followers in 30 days
• Set up payment: Stripe for international clients — payments should work before you have a product to sell
• Run a Demo Day: Zoom with 5–10 potential clients. Show three agents live. Collect pre-orders or signed commitments

Weekly challenge

Public case study + demo day + LinkedIn series started. At least three people say "I want this for my business." That's market signal.

W14 · Days 95–100 Day 100 — results Jun 10–16 0%

100 days done. Count the results, document the wins, build the plan for the next level.

• Count up: number of agents, first MRR, clients, case studies, followers. Compare with where you started on day 1
• Write a retrospective: what worked, what didn't, what you'd do differently. Publish it — that's content and personal brand at once
• Set goals for the next 100 days: first $1K MRR, five stable clients, a launched product
• Tick this checkbox. You built something real in 100 days.

March 8, 2026 · Progress saved in browser · n8n.io · OpenAI · Railway