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