Watch an agent plan dog walks

Run a demo

The plan

Street-following route from OpenRouteService; ordering by OR-Tools. The map draws from the route tool's own GeoJSON — never from the model's retelling.

How it works

The agent is a hand-rolled loop — no framework — around a small model (loading…) doing native tool calling. Three mechanisms keep it honest: a referee that validates every tool call against its schema before anything executes (violations bounce back as errors the model reads and corrects); an auditor — plain code, no AI — that refuses to accept a finished plan until every dog's actual walk interval has a weather check at that dog's location; and a structured finish line: the model ends by calling a submit_plan tool whose validated arguments are the final plan. Weather verdicts come from threshold tables in code, not from the model's judgment — the model relays them and plans around them.

The tools

• geocode_addresses — turns street addresses into coordinates (Nominatim / OpenStreetMap; batched and cached).
• check_weather — hourly forecast for a place, date, and time window (Open-Meteo), reduced in code to a deterministic verdict: OK, CAUTION, SHORTEN, or DO NOT WALK.
• optimize_route — real street-walking distances (OpenRouteService) fed to a route-ordering solver (OR-Tools), returning the visiting order, each dog's walk interval, and the map geometry.
• submit_plan — the finish line: a tool with no implementation, whose schema-validated arguments are the plan itself.

The service runs on a retired laptop in a closet (the inference is rented per-token); the whole stack costs about a dollar a month. Source, tests, and the full build history: the repo.