Why 'Recursive Language Models': a harness invoking harnesses is itself a model

a1zhang · x · 2026-11-20

Alex Zhang explains the origin of the RLM name: experiments show that harnesses recursively invoking harnesses exhibit the same locally in-distribution property, with the harness now abstracted as a sub-LLM — the harness is treated as a model.

He argues we shouldn't think of harnesses only as complex scaffolds like SWE-Agent or Claude Code; even a language model that merely invokes swarms of language models is a harness. Many ideas in the paper motivated the RLM, and he suspects many other equally or more useful harness designs remain to be found.

Related event: MIT Paper: LM Harnesses Are Compositional Generalizers, Not Crutches(10 posts)→

Original post →

More from coding & agent

coding & agent channel →