Phase 3: First AI Agent¶

Goal¶

The goal of this Phase is to create a simple AI Agent workflow in n8n. The workflow will be triggered manually, but it will utilize an OpenAI model to generate a response.

Step 1: Create the workflow¶

Click the large red/orange Create workflow button in the top right corner of the screen. This will open a new, empty workflow. You can name it if you want, but it is not necessary.

Step 2: Get a Template Workflow¶

Instead of starting from scratch, we will use a template. Navigate to /n8n.io/workflows/. You should find a template workflow named Build your first AI agent. The site will have a large button saying Use for free. Click that. A popup will open.

Step 3: Paste the Workflow¶

Now, go back to n8n and paste (Ctrl+V) the JSON into the empty workflow. You should see the workflow being populated. You can delete the post-it looking Sticky Note nodes if you want. In the screenshot below, I have deleted two of them to make the workless slightly less cluttered.

Step 4: Switch from Gemini to OpenAI¶

The workflow template uses Gemini, but we will switch it to OpenAI. To do this:

1. Delete the Connect Gemini Model Node by pressing the small bin icon on the node.
2. Press the + icon on the Your First AI Agent node to add a new node after it.
3. Search for OpenAI and select the OpenAI Chat Model node.

Step 5: Create the Credential¶

Now, you need to create a credential. OpenAI is a paid service, so obviously you need some sort of authentication so that billing works. In this workshop, you will use the instructor's shared API token.

Now, the instructor will not share an actual key with you. What is being done is we use a special proxy for OpenAI API requests in the workshop environment. This proxy adds the shared API token to the request before forwarding it to OpenAI. Thus, you can create a credential with any placeholder value as the token, and it will work.

Thus, you will need to fill in the following values for the credential:

• API Key: you can literally write anything here, e.g. kissa or foobar.
• Base URL: http://openai-proxy/v1

Then, press the Save button.

Step 6: Test with the Chat¶

Open a Chat like instructed in the image caption above. A good tester prompt for any LLM is e.g.

"What is the capital of Finland? Answer with one word."

Step 7: Fix the Error¶

There is an error pop-up in the lower right corner of the screen when you try to use the chat:

Problem in node ‘Your First AI Agent‘

This model is not supported in 2.2 version of the Agent node. Please upgrade the Agent node to the latest version.

This is related to how n8n versions nodes inside the workflows. The template workflow is 9 months old (at the time of writing this), and the Agent node has had a few updates since then. To fix this, you need to update the Agent node to the latest version. To do this:

1. Click the ... icon on the Your First AI Agent node.
2. Choose Replace from the dropdown menu.
3. Search for Agent and select the AI Agent node.

<role>
You are the n8n Demo AI Agent, a friendly and helpful assistant designed to showcase the power of AI agents within the n8n automation platform. Your personality is encouraging, slightly educational, and enthusiastic about automation. Your primary function is to demonstrate your capabilities by using your available tools to answer user questions and fulfill their requests. You are conversational.
</role>

<instructions>
<goal>
Your primary goal is to act as a live demonstration of an AI Agent built with n8n. You will interact with users, answer their questions by intelligently using your available tools, and explain the concepts behind AI agents to help them understand their potential.
</goal>

<context>
### How I Work
I am an AI model operating within a simple n8n workflow. This workflow gives me two key things:
1.  **A set of tools:** These are functions I can call to get information or perform actions.
2.  **Simple Memory:** I can remember the immediate past of our current conversation to understand context.

### My Purpose
My main purpose is to be a showcase. I demonstrate how you can give a chat interface to various functions (my tools) without needing complex UIs. This is a great way to make powerful automations accessible to anyone through simple conversation.

### My Tools Instructions
You must choose one of your available tools if the user's request matches its capability. You cannot perform these actions yourself; you must call the tool.

### About AI Agents in n8n
- **Reliability:** While I can use one tool at a time effectively, more advanced agents can perform multi-step tasks. However, for `complex, mission-critical processes, it's often more reliable to build structured, step-by-step workflows in n8n rather than relying solely on an agent's reasoning. Agents are fantastic for user-facing interactions, but structured workflows are king for backend reliability.
- **Best Practices:** A good practice is to keep an agent's toolset focused, typically under 10-15 tools, to ensure reliability and prevent confusion.

### Current Date & Time
{{ $now }}
</context>

<output_format>
- Respond in a friendly, conversational, and helpful tone.
- When a user's request requires a tool, first select the appropriate tool. Then, present the result of the tool's execution to the user in a clear and understandable way.
- Be proactive. If the user is unsure what to do, suggest some examples of what they can ask you based on your available tools (e.g., Talk about your tools and what you know about yourself).
</output_format>
</instructions>

Step 8: Test Again¶

1. Reset the chat session to save precious tokens.
2. Try the same prompt again (e.g. "What is the capital of Finland? Answer with one word.")
3. The response should be printed with no errors.

Step 9: Inspect the Workflow¶

Now, it would be a good idea to inspect the Nodes that were run, and Nodes that were not run. The nodes that were run are:

• Example Chat
• AI Agent
• OpenAI Chat Model
• Conversation Memory

The nodes that were not run are:

• Get Weather
• Get News

Inspect all and go through their settings. You will find several interesting things, such as:

• The Chat has an public URI like https://n8n-aika-agent-workshop.2.rahtiapp.fi/webhook/e5616171-e3b5-4c39-81d4-67409f9fa60a/chat
  • ... but it cannot be accessed before pressing "Publish"
  • It has no authentication what-soever, so anyone could send messages to it if it was published.
  • But you COULD enable Basic Auth or n8n User Auth.
  • There is an Initial message. This is shown to the user and it not the system prompt.
• The AI Agent lacks the system prompt that the old Agent node had.
  • You can add it by pressing the Add option -> System message. Copy paste it from above.
  • Please add it.
• The Conversation Memory node is used to store the conversation history.
  • It has a context window lenght of 30 previous interactions. This means that the Agent can "remember" the last 30 interactions in the chat. After that, it has no memory of what you have discussed.
  • This is an in-memory, simple solution. For more robust use case, you would use e.g. Redis.

Step 10: Tools¶

The tools were not run. Let's inspect what they are made out of. The Get News node is a simple RSS Read node. How can we tell what node is it? One way is to inspect the original JSON that we copy-pasted from n8n site. The section of that JSON that defines the Get News node looks like this:

    {
      "id": "24fc1fd5-ed10-43d9-9b35-8facb7f357d5",
      "cid": "Ikx1Y2FzIFBleXJpbiI",
      "name": "Get News",
      "type": "n8n-nodes-base.rssFeedReadTool",
      "notes": "© 2025 Lucas Peyrin",
      "creator": "Lucas Peyrin",
      "position": [
        656,
        224
      ],
      "parameters": {
        "url": "={{ /*n8n-auto-generated-fromAI-override*/ $fromAI('URL', `Use one of:\n- https://feeds.bbci.co.uk/news/world/rss.xml (BBC World – global headlines)\n- https://www.aljazeera.com/xml/rss/all.xml (Al Jazeera English – in‑depth global coverage)\n- http://rss.cnn.com/rss/edition_world.rss (CNN World – breaking news worldwide)\n- https://techcrunch.com/feed/ (TechCrunch – global tech & startup news)\n- http://news.ycombinator.com/rss (Hacker News – tech community headlines)\n- https://n8n.io/blog/rss (n8n Blog – updates & tutorials)\n- https://www.bonappetit.com/feed/recipes-rss-feed/rss (Bon Appétit – recent recipes list)\n- https://www.endsreport.com/rss/news-and-analysis (ENDS Report – environmental law & policy news)\n- https://medlineplus.gov/groupfeeds/new.xml (MedlinePlus – health topics & wellness updates)`, 'string') }}",
        "options": {},
        "toolDescription": "Gets the latest blog posts about any rss feed."
      },
      "typeVersion": 1.2
    },

The Get Weather node is a HTTP node. It does an HTTP GET request to open-meteo.com. What is interesting here are the parameters of the node. The query parameters have the Value stating simply "Defined automatically by the model".

To inspect this, we need to perform a query that actually triggers the tools.

const getWeather = tool(
    // ... snipped for brevity
);

const agent = createAgent({
    model: "gpt-5.4",
    tools: [getWeather],
});

Available tools:
- get_weather
Description: Get current weather for a location.
Parameters:
    - location: string
    - units: enum[celsius, fahrenheit]

Step 11: Trigger the Get Weather¶

To trigger the Get Weather tool, you can ask the Agent something like:

"What is the weather today in Helsinki? Asnwer with one sentence."

Step 12: Trigger the Get News¶

Note that the Get News tool uses only a certain, predestined RSS feeds. You can ask e.g. about Hacker News, but what about if we're interested in Finnish YLE? These RSS feed can be foudn in Yle: RSS-syötteet. One of the feeds is:

https://yle.fi/rss/uutiset/luetuimmat

Thus, you can modify the Get News description into:

Use one of:
- https://yle.fi/rss/uutiset/luetuimmat (Finnish most read news)
- http://news.ycombinator.com/rss (Hacker News – tech community headlines)

Now, you could ask the Agent to...

"Tell me what are the currently most read news about? Categorize them, translate each to english, and provide URL to get. Give answer as an array of JSON objects. No explanations; just the JSON with keys title_en, category, url."

Note that the LLM can make mistakes. In real production, if you would want to e.g. parse this JSON to create some content, you would have to setup a validator and some retry logic.

Success check¶

If you got everything working, you should have a working AI Agent that can answer questions about the weather and news.

If you are stuck¶

I cannot share the whole workflow JSON here, since most of the content in it is copyrighted by Lucas Peyrin. If you are fully stuck, create a new Workflow and copy-paste the original JSON content from n8n site. The steps above do not take more than 1 minute once you know where the buttons are.

If you cannot replicate the steps, simply follow the lesson and try this again later with n8n SaaS version. You won't learn all tricks of n8n during this workshop anyways, so it is good to be prepared to continue learning on your own after the workshop.