Glossary · Term

Agent A

← all terms

Definition

Plain language

A bare-bones AI math agent that just iterates with a language model and a proof checker.

As stated in the literature

The simplest agent variant in the DeepMind Erdős work: a Ralph-style loop pairing an LLM (Gemini 3.1 Pro) with the Lean compiler, no evolution or AlphaProof, run in parallel until a proof compiles.

Also called: Agent B, Agent C, Agent D

Why it matters: It's a useful baseline showing how much pure repetition with a strong model and a verifier can accomplish, before any sophisticated search is added.

For example, Agent A just keeps asking Gemini to write a Lean proof, checks it with the compiler, and repeats — no fancy search.

Heard on the show

“5-72B in self-play, with chain-of-thought reasoning enabled, plays Heads eighty-eight percent of the time as Agent A.”
Episode 018 — Language Models Compute the Rational Move, Then Override It

Related terms