Definition
Plain language
A way of passing credit backward down a chain, where each step pays the one before it for setting things up well.
As stated in the literature
A credit-assignment scheme from John Holland's learning classifier systems in which each activated rule pays the rule that enabled it, propagating reward backward through a chain without an explicit value function; reused in Economy of Minds via backward inter-agent payments.
Why it matters: It matters because it offers a way to reward the early moves that made a later success possible, without needing a central system to track the whole chain.
For example, when a series of actions leads to a reward, each step pays a share back to the step that set it up, so credit flows down the chain like buckets of water passed hand to hand.
Heard on the show
“This is the bucket brigade idea, right?”Episode 107 — How a Market of Crippled AI Agents Outscored One Unrestricted Model