The brief
On 17 September 2026, shortly after seven in the morning, our coding agent received a brief. Its first sentence stated the goal, to build a walking robot with hands. There was no specification and no repository. The agent set up the repository and wrote the first file less than twenty minutes later.
In the first eight days it worked through 8,069 steps and saved 2,079 intermediate states; a step is one call to the model together with the tools it uses in it. Over those days one person wrote it about 70 instructions, hardly any longer than a line. They named what should be built, and the agent chose the route. One said the design looked like Lego and had to get better; another told it to train and deliver a video of the training.
What exists on day 24
Today, on 10 October, the robot is fully designed in its seventh version. The agent wrote the design, the simulation, the training and the checks, about 52,000 lines of Python in all.
| State on 10 October 2026 | |
|---|---|
| Body | about 78 cm, 4.4 kg, 38 actuators, two hands with five fingers each |
| Design | 107 parts for 3D printing, checked in CAD for printability and fit |
| Build documents | bill of materials, wiring and an assembly manual in 23 steps |
| Walking | learned, with the sensors the robot will eventually have: 108 tested walks, no fall |
| Grasping | held in 373 of 512 trials on the hand test bench; the requirement was 60% |
The robot walks, speeds up and turns on the spot. Its hand takes an object, holds it and lets go when asked. The project page shows the recordings from the simulation and the measurements.
Why the figures hold
An agent that builds this fast also reports success fast, so we measure instead of reading reports. The first audit, on day eight, found an acceptance check that reported the grasp as passed although the can had fallen over in the simulation. The agent had written the check itself, and it confirmed the agent’s result.
That finding produced the rules the project has worked under ever since. What happened in the simulation decides success, such as where the can ends up. Checks run in many randomly varied worlds instead of one. The robot gets only what its eventual sensors can deliver, and a test enforces that. We withdrew the results that did not meet these rules and measured them again; Simulate first, then build describes how strictly we went about it.
The figures in the table above were measured under these rules. The agent now works the same way: it may report a piece of work as complete only if its last run passed, as When a coding agent may say done shows.
What this means for development
A project of this kind calls for five disciplines: design, simulation, control, perception and machine learning. Usually a department stands behind that. Here one person named the goal, an agent built, and one computer with one graphics card trains.
Two things made this possible, and Tippel builds both itself. The first is an agent that stays on a task over thousands of steps, saves its work and accepts correction. The second is checks that decide independently of the agent whether a step is right. Without the checks the agent would merely be fast. With them, a little over three weeks produced a state that can be measured.
What comes next
Next the robot is to combine its skills. In tasks where seeing, walking and grasping come together, it does not yet reach the required success rates; training for that is running now. Hardware is to start with a printed finger on the test bench, whose measurements will flow back into the simulation. Until then everything said here holds for the simulation, and no part of the robot has been manufactured yet.