Definition
Plain language
A two-agent design where one agent proposes work and another grades it.
As stated in the literature
A multi-agent pattern in which proposer agents produce candidates and verifier agents critique them, with separation of write permissions; used in RMA and other agentic reasoning systems.
Why it matters: Splitting the writer from the grader makes it harder for an agent to talk itself into a wrong answer, since approval requires passing an independent check.
For example, three proposer agents might each suggest a proof step while three verifier agents independently check the logic before anything is committed to shared memory.
Heard on the show
“The authors aren't inventing chain-of-thought reasoning here, they aren't inventing retrieval, they aren't inventing the proposer-verifier pattern.”Episode 076 — Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math