# RLM 框架：定义子智能体深度与工具调用层级

- 来源：dex (@dexhorthy)
- 发布时间：2026-07-30 11:26
- AIHOT 分数：38
- AIHOT 链接：https://aihot.virxact.com/items/cms6ygwed04kyrouq0jo33jaw
- 原文链接：https://x.com/dexhorthy/status/2082669044780695820

## AI 摘要

Dex Horthy 提出 RLM（递归语言模型）框架，将其定义为 `f(max_subagent_depth, min_subagent_depth)`。`max_subagent_depth` 表示最大子智能体层级数，`min_subagent_depth` 表示允许调用非 Agent 工具的最低层级。

## 正文

was riffing about RLMs recently… decided there are two "most interesting" dimensions

preface - my simple definition of RLM is just N-depth subagents， and wielding context well there

but… what if you frame it as：

RLM = f（max_subagent_depth， min_subagent_depth）

where max_subagent_depth is self-explanatory

and min_subagent_depth is "how many layers do you have to go before tools other than Agent（） are allowed"

so RLM（3，1） means max 3 levels and the root agent can only invoke subagents， no other tools

bu tyou could have RLM（1，0） - claude code default where the parent gets tools + one layer of subagents

or RLM（2，1） or RLM（5，2） where parent AND its subs are blocked from all tools and tool calling agents must start at layer 3

（h/t allison for the inspo）