Recursive Language Models

9 points by adsouza

adsouza

https://www.primeintellect.ai/blog/rlm is another piece about the same topic.

simonw

This piece by Drew Breunig has a useful worked example showing how this technique works: https://www.dbreunig.com/2026/02/09/the-potential-of-rlms.html

I still don't fully get how RLMs differ from pointing a Claude Code instance with grep and similar tools at a large file and telling it to analyze with subagents. Could be that RLMs can work with a smaller set of tools defined - also the subagent piece becomes a more explicit llm_query() tool.

nitinpassa

There is a definition provided in the article that makes it more clear.

The environment E abstractly determines the control flow of how the language model M is prompted, queried, and handled to provide a final output.

claude code is opinionated about how context is managed: sequential accumulation until compaction and delegation to subagents for context/task offloading. none of agents, mcp, skills, tools, or even the sdk provide a direct mechanism to change this. IIUC, RLMs hoist this harness design decision into the environment E.

with that said, you can get pretty close to an RLM style impl with some code around the sdk, and Id guess that would be good enough for a lot of use cases.