Glossary · Term

Ralph loop

← all terms

Definition

Plain language

A dead-simple way of running an AI: just let it try, show it the error, and let it try again until it works.

As stated in the literature

A minimal agent harness pattern — named after a Geoffrey Huntley blog post — that repeatedly feeds compiler or tool errors back to an LLM until output passes; the basic-agent baseline in the DeepMind Erdős work.

Why it matters: It's a surprisingly strong baseline that shows just how far you can get with no clever planning, only a tight error-feedback cycle — a useful sanity check before adding complexity.

For example, the agent writes some code, the compiler complains about a missing semicolon, the loop pastes the error back in, and the agent fixes it — over and over until the build succeeds.

Heard on the show

“The Ralph loop, named after a blog post, is just an agent attacking the same identical task over and over to iteratively improve.”
Episode 160 — Training an AI to Take Its Own Notes, So Its Future Self Works Better

Mentioned in 2 episodes

  1. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
  2. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won

Related terms