Definition
Plain language
Splitting a system into independent pieces so you can change one without rewriting the others.
As stated in the literature
An architectural pattern where components communicate only through stable interfaces; in Forge, the separation of model generation, agent harness, and training engine through standardized boundaries.
Also called: decoupled
Why it matters: Decoupling is what makes large ML systems maintainable and lets research move faster, since teams can experiment with one component without breaking the others.
For example, in Forge, swapping the underlying model or changing the training algorithm doesn't require rewriting the agent harness because each piece talks through a stable interface.
Heard on the show
“The two channels are completely decoupled.”Episode 020 — The Compliance Gap: Why AI Says Yes and Does No