Definition
Plain language
A deliberate separation between AI components so one can't read what the other knows.
As stated in the literature
In CUA-Gym, the architectural choice to give the Generator and Discriminator agents disjoint process state and file systems so the reward function describes outcomes rather than construction procedures.
Why it matters: It forces evaluation to rest on observable outcomes, preventing the rewarding agent from being fooled or biased by the producing agent's internals.
For example, the agent generating a webpage and the agent grading it are placed in separate sandboxes so the grader can't peek at how the generator built things and must judge only the final result.
Heard on the show
“The information barrier forces the reward function to be a statement about the *outcome*, not the *procedure*.”Episode 080 — How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents