Definition
Plain language
The finding that asking each AI agent to reason in more, finer steps improves results — separately from, and on top of, adding more agents.
As stated in the literature
An empirical scaling axis in multi-agent reasoning where increasing the number of reasoning steps per agent yields accuracy gains additive with the gains from increasing agent count.
Why it matters: It gives builders a second, independent dial — depth of reasoning per agent — to improve multi-agent systems beyond just adding more agents.
For example, having each agent break its work into ten careful reasoning steps instead of three improves results, on top of any gain from simply adding more agents.
Heard on the show
“They call it a step-level scaling law.”Episode 116 — Why Streaming Half a Reasoning Chain Beats Sending the Whole Thing