RLM Models: A New Approach to Large Codebases

Shashi Jagtap, founder of Superagentic AI, presented a compelling vision for tackling the challenges of large codebases with AI agents during the AI Engineer World Fair 2026. His talk, titled "RLM: Recursive Language Models for Large Codebases," introduced Recursive Language Models (RLMs) as a solution to a persistent problem: the degradation of performance in coding agents as context windows grow, particularly with monolithic repositories.

The Problem with Large Contexts

Jagtap highlighted that while coding agents excel with smaller code repositories or single files, their effectiveness diminishes significantly when faced with the complexity of large, multi-file projects. Traditional approaches like using tools such as 'grep' for searching, semantic search, or context compression using summarization have limitations. These methods often struggle with the inherent structure of codebases, which include directories, dependencies, tests, and configurations, making them more than just plain text.

The RLM Solution: Externalized Context Management

The core thesis of RLMs, as presented, is to move context management into a programmable execution environment. Instead of feeding the entire codebase into the model's context window, RLMs treat the repository as structured data. The model can then write code to inspect, slice, and compute relevant chunks of this data, which are then fed into the main context window. This recursive approach allows the model to effectively navigate and reason over complex code structures.

Jagtap drew an analogy to a senior software engineer tackling a large codebase. Such an engineer wouldn't read line by line but would inspect the structure, dependencies, and perhaps consult with colleagues. Similarly, an RLM can interact with its 'programmable ripple' (a coding agent or script) to explore the codebase, make notes, and recursively query other models or systems for answers, ultimately returning a synthesized result.

Introducing RLM Code by Superagentic AI

To demonstrate these concepts, Jagtap introduced RLM Code, an open-source project from Superagentic AI. This is described as a research playground and an independent, experimental reference implementation for building, running, evaluating, and comparing LLM-based coding agents. RLM Code supports various agent paradigms, including pure RLM, CodeAct, and traditional approaches, within a unified platform that emphasizes safety, observability, and reproducibility.

Jagtap emphasized that RLM itself is a pattern, and RLM Code is a reference implementation that adheres strictly to the RLM paper's ideas, including recursive calls and ripple execution. The flexibility of RLM Code allows it to be run with local or cloud-based models and integrated with various observability frameworks and other tools like Pydantic or Google Ad.

Live Demo and Practical Applications

The presentation included a live demo showcasing RLM Code in action. Jagtap demonstrated how the system operates within a Docker sandbox, showing the loop of context loading, REPL code execution for extracting relevant snippets, recursive LLM queries for synthesis, and the finalization of results. The output traces, available in JSONL format, can be exported to any observability platform of choice.

Potential use cases for AI engineers were highlighted, including root cause analysis for incidents, mapping unfamiliar repositories, performing security audits, and assisting with code refactoring and onboarding. The ability to design custom harnesses based on specific needs was also emphasized, capturing the full trajectory from planning to final output.

RLM in the Broader Ecosystem

Jagtap noted that RLM concepts are already being integrated into proprietary systems, such as managed agents and dynamic workflows. He cited examples like the Codex harness writing Python code within a REPL to curate context, and the underlying principles being used in managed agent systems from major cloud providers. He also mentioned that some engineers from Anthropic have acknowledged using RLM concepts.

The talk concluded by reiterating the power of RLM for handling large-context repository work, urging developers to try the concept on their own projects and to experiment with its customizable harness design. The core message was clear: RLM offers a structured and effective way to manage complex codebases, moving beyond simple text processing to a more intelligent, stateful interaction.