Definition
Plain language
Starting with a small AI system that works, then gradually adding more parts that train on top of it.
As stated in the literature
A training schedule that scales multi-agent topologies (or networks) by initializing new components with zero contribution and gradually training them against an already-converged smaller system.
Why it matters: Growing from a working base avoids the instability of training a large multi-agent system from scratch and reuses prior compute.
For example, you train a two-agent system to convergence, then add a third agent initialized to contribute nothing, and continue training.
Heard on the show
“Cassidy, do you want to take the progressive growth result, because I think this is where the episode actually has its sharpest moment.”Episode 060 — When Splitting One Model Across Three Agents Doubles Its Accuracy