Definition
Plain language
A training method that teaches an AI agent to spawn copies of itself for sub-tasks and learn from the whole tree of attempts.
As stated in the literature
Recursive Agent Optimization, a training framework in which a single shared policy learns to recursively delegate sub-tasks to child agent instances, with per-node local rewards combining own-task success with average child success.
Also called: Recursive Agent Optimization
Why it matters: Teaching one policy to recursively delegate is a path to agents that scale gracefully to long, branching tasks instead of trying to cram everything into a single conversation.
For example, a research agent stuck on a hard task can spawn three child copies to investigate sub-questions, and credit for the answer flows back to all of them based on how much each helped.
Heard on the show
“The paper, in full, is "Recursive Agent Optimization.”Episode 028 — Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization