Tiny Agents, Not AGI: Building a 5-Node Marketing-Idea Generator in n8n

tl;dr "Agents" sound intimidating, but the smallest useful one I’ve built fits neatly into a five-node n8n workflow that runs every Monday, creates fresh marketing ideas for BoxBoard, and stores them for later use.

1. Wait—what is an “agent”?

If you’ve opened LinkedIn or Product Hunt in the last year, you’ve heard

“Our platform now supports agents!”

Under the hood, most of those agents are just workflows that contain at least one step where an LLM makes a decision:

• Trigger – something kicks the flow off
  
• LLM call – the “thinking” step
  
• Action – data is stored, sent, or transformed
  

That’s it. 90 % plumbing, 10 % AI.

2. My use-case: weekly marketing inspiration

Every Monday I used to spend 20 minutes brainstorming LinkedIn ideas for BoxBoard. That’s precisely the kind of low-leverage chore automations love.

The initial 4-node proof-of-concept

1. Cron Trigger – fires at 08:00 on Mondays
   
2. OpenAI – “Give me 10 marketing post ideas for BoxBoard.”
   
3. Code – turn the JSON into rows
   
4. Postgres Insert – save into marketing_ideas
   

It worked! But after week two the model repeated itself—classic LLM amnesia.

The current 5-node version

(Screenshot below)

1. Cron Trigger
   
2. Postgres Select – pull last week’s ideas
   
3. OpenAI – “Give me 10 *new** ideas not found in this list…”*
   
4. Code – format rows
   
5. Postgres Insert – write the fresh set
   

3. Why Postgres and not Sheets / Airtable?

• Deduping & filtering – SELECT DISTINCT is trivial in SQL.
  
• Joins – later I can enrich ideas with performance metrics.
  
• Scalability – when this grows into a full “marketing squad” of flows, I already have a relational backbone.
  
• Rails front-end – I surface the table in a lightweight admin panel; no zap-to-sheet gymnastics.

Use whatever datastore fits your stack, but SQL saves future you a refactor.

4. Extending the agent

• Add a Slack node – DM myself the top idea every Monday morning.
  
• Hook into Notion – auto-create a content brief with tone & target audience filled in.
  
• Feedback loop – after a post goes live, write engagement stats back to Postgres so the next prompt can prioritize high-performing themes.

5. Takeaways

1. Small is beautiful. An “agent” can start with one LLM call and still deliver real value.
   
2. Think like a PM. Break ambitious ideas into tiny shippable tasks; wire them; iterate.
   
3. Data matters. Choosing a real database early unlocks smarter prompts and future automation layers.

Where could a five-node flow erase your Monday busywork?

Questions or war stories about your own tiny agents?
Ping me on Twitter—always happy to swap notes.