Amazon Bedrock AgentCore and Strands Agents SDK enable Recursive Language Models to process documents without context window limits

AWS introduced a method using Amazon Bedrock AgentCore Code Interpreter and the Strands Agents SDK to implement Recursive Language Models (RLMs), which treat input documents as an external environment rather than feeding them directly into a model's context window.

The approach works by having a root LLM agent write Python code to search and analyze document sections iteratively, delegating semantic analysis to sub-LLM calls via Amazon Bedrock while keeping intermediate results as Python variables in a sandboxed environment with persistent state, rather than consuming the root LLM's context window.

The architecture was evaluated against two baselines: a Base approach using a 200K token context window, and a Long Context approach using Claude's 1 million token context window, tested on the Financial Multi-Document QA subset of LongBench v2.