When a coding agent works, it spends dozens of steps reasoning, editing code, and running commands. Yet little is known about what the underlying model internally encodes about the programs it works on.
We looked inside the model's hidden states to see what they actually encode.
