A retry loop cannot fix the toolchain

An early agent: give it a blueprint, it generates an Android project, runs the build, and feeds errors back to the model for up to three tries.

Over one day it ran eleven times. One run built, on its second attempt. Four used all three attempts and got nowhere. Six stopped on the first attempt, because the problem was not the code.

The model had been asked to repair a machine it could not see. Handed a broken toolchain, it will edit a perfectly good file with total confidence.

So the agent learned to sort failures before retrying them: environment or code. Environment errors stop the loop.

The project was abandoned after three commits. The sorting survived it.