RLM (Recursive Language Models) — Data Scientist Agent Embedded in Programs via DSPy SandboxSerializable

Kevin Madura (@kmad) shares his March 22 2026 blog post introducing a new pattern for embedding LLM-based data analysis agents directly into programs using Recursive Language Models (RLMs). The approach extends DSPy's SandboxSerializable protocol to expose DataFrames to a persistent REPL-based LLM loop — the model iterates, writes code, inspects results, and recurses until analysis is complete. Benchmarked on DABench (257 questions, 68 CSVs), Qwen 3.5 397B reaches 86.8% accuracy with 2.8 average iterations. The upstream library (alexzhang13/rlm, 3177 stars, MIT, Python) provides plug-and-play inference with Docker/Modal sandbox support.