Recursive Agent Harness Enhances Long-Context Reasoning in AI Coding Agents

Recursive language models (RLMs) have demonstrated that recursion over model calls is a potent technique for long-context reasoning. Recent developments in production coding agents, particularly Anthropic's dynamic workflows, have led to the introduction of the Recursive Agent Harness (RAH), which employs a full agent framework equipped with filesystem tools, code execution, and planning capabilities. This study evaluates RAH's performance against existing baselines, revealing that it improved the Codex coding-agent baseline accuracy from 71.75% to 81.36% on the Oolong-Synthetic dataset. Additionally, with the Claude Sonnet 4.5 backbone, RAH achieved an accuracy of 89.77%, highlighting the effectiveness of harness recursion in enhancing coding agent performance.