Sign InOpen Brain
AI EngineerVideoSource Linked

It’s Tokens All The Way Down: How RLMs are Different — Kevin Madura, AlixPartners

RLMs keep large inputs in a programmable environment, letting a model inspect them with code and delegate subsets to submodels instead of attending to every token at once.

AI Engineer · Sep 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

A recursive language model treats context as a variable inside a Python-style REPL, writes code against it, and can delegate selected work to another model. The talk contrasts this with asking a base model to find **12 numbers across 30,000 tokens** directly.

Practical Implication

Consider the pattern for large logs, data frames, repositories, or other inputs that benefit from deterministic slicing and computation. Define typed inputs and outputs, cap iterations, and return only relevant intermediate results to the main model.

Agent-Ready Context
A recursive language model treats context as a variable inside a Python-style REPL, writes code against it, and can delegate selected work to another model. The talk contrasts this with asking a base model to find **12 numbers across 30,000 tokens** directly.

Consider the pattern for large logs, data frames, repositories, or other inputs that benefit from deterministic slicing and computation. Define typed inputs and outputs, cap iterations, and return only relevant intermediate results to the main model.

The cited benchmark rose from **2.6% to 45.4% accuracy**, but the coding-agent comparisons were described as preliminary and potentially unfair. Cost, stopping behavior, sandboxing, and benefits over a carefully built conventional agent still require task-specific tests.
Connected Context · Feed7 Judgment

This introduces recursive model calls as a selective computation primitive rather than a general multi-agent topology: keep large context outside the main prompt, inspect and slice it deterministically, and delegate only bounded fragments. It strengthens the case for typed interfaces, iteration limits, and compact intermediate results, but the benchmark does not yet show that this outperforms a carefully designed conventional coding-agent harness on comparable tasks.

Why We Killed Our Multi-Agent Pipeline — Subbiah Sethuraman and Abhilash Asokan, ZS AssociatesThe failed specialist chain favors one owner with bounded investigations; the recursive pattern offers a compatible way to delegate selected fragments without distributing end-to-end ownership.Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC BerkeleyIts guidance on narrow subagent scope, isolated outputs, stop reasons, and compaction directly supplies controls for the recursive loop’s delegation, context return, and termination risks.Agent swarms and the new model economicsBoth separate planning context from worker computation to reduce coordination burden, but the recursive model is oriented toward programmatic slicing of one large input rather than a general worker swarm.CausalForge: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal InferenceCausalForge reinforces offloading structured derivation to deterministic machinery, while also warning that verified computation does not establish that the selected task or claim is the intended one.
Context Map
agentcodingdata#reasoning#subagents#harness-engineering
Uncertainty
The cited benchmark rose from **2.6% to 45.4% accuracy**, but the coding-agent comparisons were described as preliminary and potentially unfair. Cost, stopping behavior, sandboxing, and benefits over a carefully built conventional agent still require task-specific tests.