Recursive Language Models (RLMs) are a scaffolding architecture that fundamentally differs from ReAct, CodeAct, and vanilla subagent patterns. Instead of loading the full user prompt into the LLM's context window, RLMs give the model access to a persistent Python REPL where the prompt lives as a variable. The LLM deliberately reads slices of context, writes code to transform data, and recursively invokes subagents via llm_query(). Crucially, subagent outputs are returned as Python variables in the REPL rather than being auto-loaded into the parent's context window, enabling composable answers without token-by-token regeneration. This approach enables arbitrarily long outputs, focused attention on relevant context sections, parallel subagent execution, and significant cost savings through KV cache reuse. The article walks through a concrete toy problem (counting letter 'r' in fruit/country/animal names) to illustrate how each architecture handles it, culminating in how RLMs solve it elegantly through recursive composition.
