I gave an agent a lakehouse and control of my garden

A small, two-square-metre vegetable bed of cabbage, broccoli, lettuce and spinach becomes an experiment in agent-run gardening. After relying on Claude for planting and care advice, the gardener equips the space with eleven wireless sensors measuring soil moisture and temperature, leaf wetness, weather and other conditions, then maps their locations so an agent can reason about differences across the garden. A camera adds visual monitoring - and reveals that a neighbour’s cat uses the bed as a toilet. The equipment costs about €300, far more than the vegetables are likely to be worth, prompting the joke of a €60 lettuce.

The project’s central problem is turning those measurements and images into useful decisions. The gardener builds LakehouseBox, a simple, European-hosted lakehouse where an Ecowitt gateway sends readings directly into Apache Iceberg tables, without an intermediary service. Since September 18, the system has accumulated more than four million readings in a table occupying 3.22 MB. An agent can query the history to assess drying, watering and differences between locations, rather than relying on dashboards or manually writing SQL. Because the camera only supports encrypted FTP uploads, LakehouseBox is extended with a write-only drop box that stores its images and indexes them in a queryable table. The larger idea is to let an agent combine sensor data and pictures, analyze the garden, and eventually communicate with the owner and control irrigation.