Recursive Language Models are fascinating as they manage information dynamically instead of doing a single-pass reading. Take Prime Agent for instance: it is a self improving RLM harness for coding and long running autonomous tasks. On ARC-AGI-3, it scores 95.5%, which is higher than human experts. Since these types of models treat context as a variable, their IPython kernel can program over history, call tools and launch sub agents.
Introducing Prime Agent:
A self-improving RLM harness for coding and long-running autonomous tasks.
Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state. pic.twitter.com/Bwj7q9Virh
— Prime Intellect (@PrimeIntellect) August 5, 2026
This is a coding agent but can also be used for other tasks. It can for example run “on challenging long-horizon settings like Factorio and MazeBench.”
[HT]
