Building a dog-walking agent that runs on my own machine

I wanted a small, honest test of “agentic” AI — something with a real decision at the end, not a chatbot that trails off. So I built an agent that answers one question: should I walk the dog right now?

Make it decide, not chat

The first choice that mattered was the output. It would’ve been easy to let the model ramble a paragraph of weather vibes. Instead I gave it a schema — a Pydantic WalkDecision with a recommendation (WALK, SHORT_WALK, WAIT, SKIP), an optional duration, and explicit reasons and risks. Forcing structure did two things: it made the answer actionable (a program could consume it), and it quietly made the model reason better — you can’t fill in risks: [] honestly without actually checking the conditions.

Give it tools, not trivia

The model doesn’t know the weather in Atlanta, and it shouldn’t pretend to. So it gets tools: geocode a city, fetch live weather and air quality (Open-Meteo), look up the daylight window, check the time. The agent decides which ones it needs for a given question — ask a simple question and it skips the API calls; ask about now and it goes and gets the real numbers. That orchestration is the actually-interesting part: the model as a planner, the tools as its hands.

Keep it local

I ran the whole thing on a local LLM through Ollama — no cloud endpoint, no API key, nothing leaving the machine. Partly privacy, partly cost, partly because I wanted to know how far a capable local model could get on a genuinely multi-step task. The answer: further than I expected. Tool-use and structured output both held up.

What I’d do next

The memory is currently in-process, so it forgets between runs — persisting facts to disk is the obvious next step, and it’s where this stops being a demo and starts being mine. A small CLI instead of a hardcoded prompt would help too.

But the core lesson stuck: a clear decision, a strict output shape, and a handful of real tools turn “an LLM” into something that feels a lot more like an engineer’s instrument than a toy.

Project notes

This project is based off the principals from Victor Dibia’s “Designing Multi-Agent Systems”, 2025.