The model writes, the code computes

The Week 1 theme, “Touch Grass”, asked for a project that takes people away from the screen and outdoors. I built autumn-walks, an AgentA program that uses a language model to decide what to do (search, call a tool, answer, refuse), in a loop, with stopping rules.See in the glossary → that sends me an autumn walk through Paris on Telegram every morning. The walk is built from the city’s trees and the day’s weather.

The concept

In an agent, the model handles what takes language. Anything with only one right answer, like distances or the choice of stops, is computed by the code.

What I did

I filtered the “Les arbres” dataset from Paris Data (220,174 rows) into a SQLite database. That leaves 2,445 autumn trees in 5 arrondissements.
I built a Strands agent with Gemma (gemma4:e2b), running locally with Ollama. It has two tools, weather and trees. Without a System promptThe fixed instructions given to the model before any conversation. A rule there reduces a behavior without eliminating it, and a prompt that's too long is followed less well.See in the glossary →, Gemma called no tool at all.
When Gemma chose the stops, the loops were 5 km long and some places didn’t exist. The code now chooses the stops, and Gemma writes the message.
The agent keeps a memory of past walks, and the streets from the last two are excluded. The message goes out on Telegram every morning at 8:30.
The code was written with Claude Code.

What I keep

The system prompt is part of the tool.
Geometry is not a language task.
Real-world use should be planned mid-week, not at the end.

The defense question

Why doesn’t Gemma choose the trees itself?

When it did, the loops were 5 km long and some places didn’t exist. The code guarantees real stops within 4 km, and Gemma handles the text, which it does well.

Submitted on October 10. I couldn’t do the walks before the deadline, and the post says so: the bot works, but it hasn’t been used outdoors yet.