Recursive Language Models: Why code over prompt beats context stuffing

CShorten30 · x · 2026-08-27

The Weaviate Podcast explores Recursive Language Models. Unlike typical agent harnesses that treat the prompt as a container to fill up, this approach treats it as a variable for the program. Alex explains that stacking tool outputs into a growing context is badly out-of-distribution for pretrained models, and that writing code over the prompt offers a more robust solution.

Original post →

More from coding & agent

coding & agent channel →