AI walkie-talkie author sculpted the case from clay because the agent can't feel the hand

There is a class of tasks where a coding agent confidently calculates everything that can be calculated, and still loses to a lump of clay. Ergonomics is exactly that kind of task. The author of a pocket walkie-talkie for a personal agent system went down this path in practice and described it on Habr.

Why a walkie-talkie at all

The idea is simple: pick up the device, squeeze the trigger, say a thought, release. The recording travels over Bluetooth to the phone (no need to take it out of your pocket), and from there the agents in a system the author calls the "Bureau", his second brain, process it. No messenger buttons and no open apps.

The interface is reduced to a minimum. All text has been removed from the screen, leaving only animation: while you hold the trigger, it records; after you release it, the recording flies off; and when the system has accepted the recording, a flash and a checkmark appear. Each event has its own sound.

There are two devices. The first version is built on the M5Stack Cardputer-Adv: ESP32-S3, a 1.14-inch screen, a microphone, and a 1500 mAh battery. The case slips on top, and a large side trigger presses the built-in button. The second version is built on the M5StickS3: the same internals, but without a keyboard and with a 250 mAh battery. By the author's estimate, such a device will last about a week on a single charge.

Where the agent hits a ceiling

The author designed the first version's case together with an agent directly in code: shapes via SDF and boolean operations, a printed trigger spring, renders. The mechanics worked out, and the neural network calculates them excellently. But in the hand, the case was awkward.

According to the author, the model understands anatomy poorly, and calculation doesn't help here. I would add that "comfortable" has no formula that can be driven into code. The feel of a palm can only be checked by a palm. So the author divided the responsibilities: the hand defines the shape, and the agent does everything else around it.

How the clay pipeline works

Everything needed for sculpting was printed on a 3D printer. The whole kit took 3.1 hours to print and used 98 g of PLA, and the dummy for the StickS3 added another 28 minutes.

• The dummy is a copy of the device enlarged by 2 mm on the sides and back. The surface has 0.8 mm dimples so the clay holds better. At the location of every button and connector there is a hole for a mark made with a black marker pin: after sculpting, you can see where the button is under the clay.
• The shooting stand is a column with a keyed peg so that the dummy sits the same way every time, plus three hemispheres at the corners. They let the scan snap into the device's coordinates and scale.
• The sculpting tools are two double-sided rakes, two smooth scrapers, and two kidney-shaped profile tools. No suitable ready-made ones were found, so the author printed his own.

Then comes the manual work: architectural clay is wrapped around the dummy so that the marks stick out, the excess is cut off with a profile tool, and the shape is constantly tried in the fist. The finished shape is placed on the stand and scanned with photogrammetry in RealityScan. Then a neural network builds a smoothed model from the scan, and the agent uses the marks to place the mechanics inside it. That is what will be printed.

What to take away from this

The main idea is broader than the walkie-talkie: if a task has a measurable part and a "felt" part, they should be separated. Give the model constraints, tolerances and component fit; give the human whatever is checked by the body. Fiducial marks and the stand link the two: the human result becomes an artifact the agent can read.

There are limitations too. The shape has so far been sculpted for one person's hand, and the author himself says he will work on making it more universal. The final printed model is not yet shown in the text. So whether the "scan → smoothed model → mechanics by marks" stage works as smoothly as intended is not yet known.

For those building hardware with agents, the takeaway is practical: don't try to squeeze ergonomics out of a prompt. Print a dummy, sculpt the shape, scan it, and give it to the agent as input data.