Today I want to share a real-world case of automating weekly reports for one of my clients. The task was standard—collect data from three sources, structure it, and send it to Slack. But the implementation approaches offered interesting tradeoffs.
Approach 1: Classic Orchestration
The first version used n8n as the control center, supplemented by custom scripts for complex parsing. Advantages:
- Complete transparency at every step
- Option for manual intervention at any point
- Low-cost GPT-3.5 queries for data cleaning
-
- Data collector (using Playwright)
- Validator (Mistral 7B)
- Report generator (GPT-4 Turbo)
-
Similar to Databricks' AI strategy, the key was responsibility separation. But there's a catch—GPT-4 tokens devour budgets when processing messy data.
The Verdict?
A hybrid solution. Simple tasks stayed in n8n, complex logic moved to agents. Lesson: full autonomy isn't ready yet, but targeted agentic workflows deliver 30% faster processing.
P.S. The coffee in Tallinn was exceptional today—maybe that's why I'm writing so much.
But it required writing tons of exception-handling code. Like the old plumber joke—"$20 for hitting the pipe, $980 for knowing where to hit."
Approach 2: Agentic Workflow
I rebuilt the system using autonomous agents. The new architecture became a chain of three microservices: