You Can't Ctrl-Z a Robot: Coding Agents in Embedded Systems

Coding agents work well in the terminal because feedback there costs nothing. Compiler errors, failing tests, stack traces, all of it arrives instantly and in writing, and the agent reads the message and tries again. As soon as code has to move something in the physical world, that stops working. Reality writes no error messages, and it has never had an undo.

Embedded engineers dealt with this long before anyone talked about agents. They build simulators to fail in, they measure what actually happened instead of what was intended, and no change reaches the hardware without proof. We put a coding agent into that same loop and look at where it earns autonomy, where it has to be stopped, and why teams whose code never touches hardware run into the same problem with every irreversible action they automate.