Skip to main content
Recursive language model (rlm(...)) spawns real child agents (SubagentRuntimeHost, RlmSubagentRuntime). Subagents run in parallel or background and return results programmatically. The RLM harness (/refine) uses trajectory evidence to update supplemental state; it does not modify the base prompt. Related: Coding agent index, Refinement
Last modified on August 17, 2026