Over the past couple of years, I’ve done a lot with ChatGPT, including building custom GPTs for better prompts, generating alt text for images, writing cold sales emails, and even creating bedtime stories. I’ve also learned other AI tools along the way, including Perplexity (which I now use more than Google) and NotebookLM.
To advance my skills, I want to begin working with AI agents. I follow many AI YouTubers, but these two had simple projects that I could build with n8n.
Jeff Su Tutorial
According to Jeff, an AI agent needs 3 things to function properly:
- A brain that’s made of a chat model (like ChatGPT) and memory
- Tools like Slack, Google Sheets, or Notion so the agent can actually do stuff
- A brain stem in the form of a system prompt
This was the first project I’ve done where I’ve needed the ChatGPT AI key, which Jeff and ChatGPT made very easy to find. Also got to use JSON files for the first time.
The AI Agent
This is the n8n workflow.
My chat input into n8n.
This is what was added into the spreadsheet. The agent worked!
I then added several more subscriptions to further test the agent.
I then added a duplicate subscription on purpose and the agent caught it.
Grace Leung Tutorial
In the Jeff Su tutorial, connecting to Google Sheets was simple (just log in with Google credentials). But in this tutorial to access Google Docs and Drive, I had to get an API key via Google Cloud. It was challenging, but ChatGPT was a huge help.
According to Grace, every AI agent needs the following:
- Role and Responsibilities: A precise, clear summary of the agent’s purpose.
- Tools and Data Access: A list of all external tools and data sources the agent can use.
- Input and Triggers: The types of user input that activate the agent’s workflow.
- Workflow Logic: The step-by-step process the agent follows to complete a task.
- Important Guardrails: The absolute boundaries and rules the agent must never violate.
- Output Format and Success Criteria: The required structure of the agent’s final response and what defines a successful outcome.
Grace’s project was a competitor analysis, which I didn’t care to do. I wanted to do it my own way, so I wrote an essay about symbolism in the film, The Godfather. The purpose of this agent would be to rewrite my essay and add primary sources to support my arguments.
The AI Agent
This is the n8n workflow.
This is my original essay with zero cited sources:
This was completely generated by Gemini.
I prompted the n8n automation and this is what it looked like after it ran.
This is the new essay with primary sources cited:
The superscripts were hard to see so I circled them in red.
I told it to add a Works Cited Section at the bottom.
The rest of Grace’s tutorial includes great instructions for creating a Visualization Agent and a Management Agent, but those features require a paid plan. I’m currently on a free trial, so I don’t have access to them.
Tina Huang Tutorial
This was the hardest project I have worked on, and I almost gave up several times. Tina’s instructions were clear, but for some reason, the AI agent would not work. It might have been because she was using an older version of n8n, the free trial instead of the paid version, text limits, or small differences in our settings.
Whenever I ran into a problem, I asked ChatGPT for help. I shared screenshots and sometimes the JSON file and explained what was not working. In the end, we added a few extra nodes and some code, and it finally worked.
The AI Agent
Click to see larger version
This agent takes a topic and a time range entered through a form. For example, I can input a topic like the Gaza War, government shutdown, or NFL stats, and specify a time range such as last week or the whole year. The agent then emails me an audio file explaining the topic within that date range, usually around four minutes long.
According to Tina, an AI agent has six core components:
-
Model: The brain that powers the agent. It can be ChatGPT, Gemini, or Claude.
-
Tools: What is needed to perform the requested task, such as access to your calendar to book appointments.
-
Knowledge and Memory: Allows the agent to remember current and previous requests.
-
Audio and Speech: Enables the creation of audio for your agent to communicate with humans.
-
Guardrails: Safety mechanisms that ensure proper behavior, such as preventing profanity or hate speech.
-
Orchestration: The systems that monitor your agent so you can continue improving it.
Trigger
This is the form that starts the automation:
Input topic and date range for how far you want to go back.
Model, Memory, & Tool
The model used for this agent is ChatGPT. The memory component is n8n’s Simple Memory. The tool used is Perplexity, connected through an API key.
The prompt, created in ChatGPT, is placed inside the AI Agent node.
Audio & Speech
This node creates an audio file from the information it receives and emails it to me.
This is the downloaded audio file. Press play to listen to it here.
Guardrails
This part of the agent checks for inappropriate content, such as abusive language. Since the information comes from Perplexity, it’s unlikely to contain anything like that. However, if it ever did, the agent would email me an alert.
Orchestration
This is where the transcript of the audio file, along with the topic and date range, is stored. This allows me to see my results visually and make changes to continuously improve the automation. The automation also evaluates the quality of response by giving it a score from 1-5.
