AI automation needs guardrails more than clever decisions
TomatoFarm is a real indoor farm growing six tomato plants. It uses sensors, cameras, pumps, mostly 3D-printed parts, a .NET backend, TimescaleDB, and a Blazor interface. Four , Claude, ChatGPT, Gemini, and Mistral, read the same farm status and vote on what should happen next.
The AI cannot control hardware directly; it can only suggest actions. Any action such as watering must pass fixed limits, cooldowns, tank-level checks, and fail-safe rules before it can run. Early question-answering failed badly because the AI invented features the farm did not have, including hydroponics, nutrient dosing, EC values, and even a wrong plant name.
The fix was grounding: answers can only use real facts from the database and . The system keeps running if one or two models fail, can assign a new lead model, and has two separate AI operators in watching data flow, hardware state, database values, and scheduled checks for problems.
Key points
- The AI suggests actions but does not directly control hardware.
- Every action must pass fixed limits, cooldowns, tank checks, and fail-safe rules.
- The system only lets AI answer from database facts and live .
- The farm can keep running even if one or two fail.
- Two separate AI operators watch the system and raise alerts when something looks wrong.