PrimeIntellect Releases Prime-Agent: A Self-Improving RLM Coding Agent
PrimeIntellect has open-sourced Prime-Agent, a self-improving agent designed for coding workflows and long-running autonomous tasks. Built on TypeScript, it utilizes Recursive Language Models (RLM) and a Continual Harness to improve its performance over time. This project addresses context window limitations by programmatically delegating tasks to subagents, representing a shift toward more scalable and autonomous AI software engineering. It also demonstrates state-of-the-art capability, achieving a 95.5% score on the ARC-AGI-3 benchmark when paired with Opus 5. Prime-Agent relies on two core abstractions: the Recursive Language Model (RLM) and the Continual Harness. By using RLMs, the agent avoids stuffing entire contexts into a single window, instead writing code that dispatches subagents over smaller context chunks.
## BACKGROUND
Traditional AI agents often struggle with long-running tasks due to context window limits, where older information is forgotten or diluted (known as context rot). Recursive Language Models (RLMs) solve this by allowing the primary agent to dynamically spawn subagents to process specific chunks of data, keeping the context focused and manageable.