Google Open-Sources Harness²: Enabling AI Agents to Self-Optimize Execution Scaffolds at Runtime

The Google Research team has open-sourced Harness², a Recursive Self-Improvement (RSI) framework. This system allows large language models to autonomously modify their external code, system prompts, toolchains, and skill libraries during task execution, aiming to address the challenge of effective self-optimization in real-world open environments where ground truth is often unavailable.

Harness² employs a dual-layer recursive architecture. The first layer, task-level recursion, utilizes a 'Propose-Probe-Compose' mechanism to evaluate scaffold modifications via behavioral contrast without external judges. The second layer, improvement-level recursion, uses a reflection module to distill debugging records into an experience library, enabling the system to retrieve effective components from past tasks for continuous learning across new assignments.

According to Google's internal benchmarks, on the WorkBuddy-Bench office productivity benchmark, the GLM-5 model equipped with Harness² saw its composite score rise from 64.2 to 75.7, an increase of 11.5 points, surpassing the native Gemini 3.5 Flash (72.4 points) which did not use the framework. Additionally, the framework delivered performance gains of 4.5 points on JobBench and 11.8 percentage points on the LAB legal benchmark. The project currently supports OpenCode and Codex environments.