Opus 5.5 Designs a Self-Watering Herb Pot on Its Own

Picture an ordinary supermarket herb pot: basil inside, drooping as soon as the water runs out. Bastian Hildner chose an unusual way to redesign it from scratch. He placed a 12-point requirements brief in front of an AI model called Opus 5.5 and typed a single sentence: implement the brief. The model asked four questions, then worked alone for nearly two hours. What surprised me most was not the speed but the sight of one agent shuttling back and forth between two separate expert programs on its own initiative.

The first half of the job happened on the electronics side. Inside KiCad, the agent drew the circuit schematic, placed the components and routed the copper traces. Both the electrical rule check and the physical rule check passed with zero errors on the first run, and the parts list plus the Gerber files for the board manufacturer came out of the same session. I was not surprised by this part: the Claude team describes in its own blog how the model reads documentation and produces reference circuits. The picture gets even more concrete on GitHub: an MCP server written for KiCad opens schematics and layouts to the agent in a token-efficient format, runs the automatic router, passes every check and closes the full write loop through 32 dedicated tools.

From circuit to enclosure: two programs, one carrier

Once the board was done, its three-dimensional model moved over to Fusion and the agent built the enclosure around the board piece by piece. About an hour later the first delivery was ready, and the cross-section view showed the tank, the float and the board all sitting where they belong. Then came four rounds of revisions: first a hose and cables between the pump and the board, then a plant to show how the pot will look on a windowsill, then a modernized enclosure with the board switching to surface-mounted parts, and finally a transparent cutaway to reveal the inner layout. Autodesk officially backs this workflow: the MCP path opened for Fusion is meant to let AI do real work inside the design environment. A community server on GitHub shows the kitchen side of it: a Fusion add-in receives commands over a TCP bridge and executes them safely on the main thread.

The connection layout here deserves a close look, because that is where the trick lives. The KiCad side has its own MCP connection and so does the Fusion side, yet the two programs exchange not a single word with each other. The agent hands the order to one side, carries the finished board model to the other and checks each result against the brief. So there is no single clever program in the middle; there are programs that each do their own job plus an agent running between them. And the last word stays with the human: the person decides what gets built, which revision to request and how much autonomy the agent enjoys in each department.

Now to the pot itself. The water tank sits at the very bottom, and its roughly one-and-a-half-liter volume keeps the plant happy for about a week. A small pump inside the tank pushes water up through a hose to a spike stuck in the soil, while a moisture sensor in the soil decides when it is time to water. The electronics rest on top, under the lid, safely separated from the water and powered through a USB-C port. A lamp on the lid shows whether all is well, a button waters on press, and when the tank runs dry the float reports back: the lamp turns red, a phone notification lands and the pump stops so it never runs dry. A study published with Atlantis Press confirms the same logic: pots that track soil moisture and water on demand follow the very same principle of steady growth with little water.

Cost, promo clip and the virtual firm

The time and cost side was disclosed too: two and a half hours of work in total, four revision rounds and 25 units of usage billed from the subscription. Hildner says he did not draw a single line himself in all that time. The bonus was a promo clip: photorealistic frames rendered from Fusion views, then a short advert assembled through the Higgsfield connection. Higgsfield is tailor-made for this kind of job, offering more than 30 image and video models through one connection, 4K output and clips of up to 15 seconds. The design files were shared openly with the Skool community, so anyone can print the pot on their own machine.

The closing section opens up a separate project: Dark Company. Hildner wants to set up a virtual firm there and connect its departments to automation one by one, deciding up front how autonomously the agent may act in each unit, from development and support to manufacturing and shipping. The thesis is bold but grounded: the goal is not a fully mechanized company but faster progress on the slow jobs where small firms struggle to hire, with AI assistance doing the lifting. Email replies, stock records and shipment tracking are the first candidates. My reading is this: the herb pot experiment was the first rehearsal of that bigger plan, and the rehearsal showed that supervised autonomy genuinely works.