ICANEWS

Engineering & Technology

Recursive Agent Harness (RAH) for Long-Context Reasoning in Coding Agents

The Recursive Agent Harness (RAH) constitutes a framework that applies recursion over complete agent harnesses, distinguishing it from the model recursion emplo

Overview

The Recursive Agent Harness (RAH) constitutes a framework that applies recursion over complete agent harnesses, distinguishing it from the model recursion employed in Recursive Language Models (RLMs). RAH frames this approach as harness recursion, presented as a code-first extension to the model recursion described in RLMs. This pattern has been observed in production coding agents, notably in Anthropic's dynamic workflows, where code is generated to spawn subagents.

Within the RAH framework, a parent agent is designed to generate and execute a script. This script facilitates the parallel spawning of subagent harnesses for handling fine-grained workloads. Concurrently, the RAH utilizes structured function calls to manage smaller subtasks.

Research Context

Recursive Language Models (RLMs) previously demonstrated the effectiveness of recursion applied to model calls for long-context reasoning tasks. Separately, production coding agents have begun to integrate code that instantiates subagents, a feature present in systems like Anthropic's dynamic workflows. The present work identifies and investigates the pattern common to these two lines of development. It characterizes this pattern as one where the recursive unit functions as a comprehensive agent harness, complete with filesystem tools, code execution capabilities, and planning functionalities, rather than merely a model call lacking such tools.

Approach

The study focused on a controlled evaluation of the Recursive Agent Harness (RAH) in the context of long-context reasoning. The experimental setup maintained the backbone model at GPT-5. This specific choice allowed for direct comparison with previously published baselines for Codex and RLMs. The evaluation aimed to isolate the contribution of the harness design from the underlying language model's capabilities.

Findings

  • When the backbone model was fixed at GPT-5, mirroring the published baselines for Codex and RLMs, the RAH design improved the Codex coding-agent baseline from 71.75% to 81.36%. This performance increase was observed on the Oolong-Synthetic dataset, which comprises 199 samples distributed across 13 context-length buckets, extending up to 4 million tokens. This gain is attributed to the harness itself rather than the core model.
  • The same RAH design, when integrated with a more capable backbone model, Claude Sonnet 4.5, achieved a performance of 89.77% on the Oolong-Synthetic dataset.

Why This Matters

The observed improvement in performance, specifically the jump from 71.75% to 81.36% on the Oolong-Synthetic benchmark with the GPT-5 backbone, suggests an efficiency in managing long-context reasoning through the Recursive Agent Harness. This indicates that the architectural design of the harness itself, rather than solely the underlying language model, contributes to enhanced performance in complex coding agent tasks, particularly those requiring extensive contextual understanding.

Key Findings