How to use RLMs in Deep Agents

Short summary

Sydney Runkle from LangChain walks through Recursive Language Models (RLMs)—a pattern where agents can call themselves to tackle large-scale tasks that break standard agents. She demos the pattern using dcode and benchmarks RLM-enabled deep agents against the Oolong dataset, showing reliable performance at 128k token context lengths where plain agents fail. The pattern enables deterministic, finite-context task execution across parallel subagents.

• •RLMs let agents call themselves recursively to handle large-scale tasks and long contexts
• •dcode demo shows a live haiku tournament across 16 parallel subagents
• •Benchmarks show RLM agents outperform plain agents at 128k token contexts on Oolong dataset

Generated with AI, which can make mistakes.