Reproduce “Recursive Agent Harnesses” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Elias Lumer, Sahil Sen, Kevin Paul, Vamse Kumar Subbiah

PricewaterhouseCoopers researchers introduce the Recursive Agent Harness (RAH), an architectural pattern extending recursive language models by making the full agent harness, equipped with tools and code execution, the fundamental recursive unit. This system improves performance on long-context, fine-grained reasoning tasks by 9.61 percentage points over prior coding agent baselines on the Oolong-Synthetic benchmark.

Key insights

  • Formally defining and evaluating "harness recursion" demonstrates a novel architectural paradigm where robust, tool-augmented agents can be recursively applied to subproblems, bridging the gap between coding agents and bare recursive language models.
  • The
  • code-first extension
  • to model recursion effectively enables per-entry LLM reasoning at scales far exceeding a single context window or fixed function-call budgets, providing a scalable solution for complex information aggregation.
  • The RAH pattern aligns with emerging industry practices in production systems, suggesting a practical and effective strategy for orchestrating agents for tasks that demand extensive context and fine-grained processing.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.