Google Open-Sources Harness², a Recursive Framework Designed to Improve AI Agents Without Retraining Models

Google Cloud AI Research has open-sourced Harness², a recursive agent optimization framework designed to improve artificial intelligence systems by refining the software and instructions surrounding an existing AI model. The approach allows developers to enhance agent performance without changing model weights or conducting additional model training.

Released as version 0.1.0, Harness² focuses on optimizing the components that determine how AI agents plan tasks, use tools, manage memory, and execute workflows.

What Is Google’s Harness² Framework?

Harness² is an open-source framework that improves an AI agent’s operating environment, commonly known as its harness.

Rather than modifying the underlying language model, the framework adjusts the surrounding infrastructure responsible for agent behavior.

The system contains eight editable components:

• System prompts

• Guardrails

• Memory

• Tool guidance

• Subagents

• Skills

• Scripts

• Plugins and middleware

These components influence how an agent interprets instructions, accesses tools, delegates responsibilities, and completes complex tasks.

Developed by researchers associated with Google Cloud AI Research and Northwestern University, Harness² is intended for real-world environments where agents encounter unfamiliar tasks without predefined answers or verification systems.

Why Harness Optimization Matters

Traditional AI improvement methods frequently depend on model retraining, labeled datasets, or established evaluation procedures.

Harness² takes a different approach by examining an agent’s behavior during task execution and identifying improvements to its supporting infrastructure.

This could help organizations improve existing AI deployments without investing in additional model training.

How Recursive Agent Harnessing Works

Harness² uses two interconnected optimization processes: task-level recursion and harnessing-level recursion.

Task-Level Optimization

The first process follows a three-stage procedure called Propose, Probe, Compose.

An agent initially completes a task using its original harness. An editor then generates two potential improvements.

The first, called an invariant candidate, addresses general weaknesses such as planning, reasoning, or tool usage.

The second, called a domain candidate, focuses on specialized knowledge and procedures relevant to the task.

Both candidates are evaluated against the original execution. A composition stage compares their behavior and retains changes supported by observed results.

Learning Across Multiple Tasks

The second process occurs after task completion.

A reflection mechanism stores successful harness configurations alongside lessons documented in an EXPERIENCE.md file.

These records help guide future optimization proposals, although each new task begins with the original base harness.

The framework supports parallel and sequential refinement strategies, controlled by a configurable search budget.

Harness² Benchmark Results Show Performance Improvements

Researchers evaluated Harness² using three professional AI benchmarks: WorkBuddy-Bench, JobBench, and Harvey’s LAB.

Testing involved Gemini 3.5 Flash and GLM-5 operating through OpenCode and Codex environments. Claude Opus 4.8 served as the default editor.

Results were averaged across three runs.

Performance After One Refinement Step

Source: Harness² research paper, Tables 1 and 2.

All four configurations improved across the three benchmarks after a single refinement step.

The largest WorkBuddy-Bench improvement occurred with GLM-5 running under OpenCode, increasing 11.5 points from 64.2 to 75.7.

The smallest reported gain was GLM-5 under Codex on JobBench, which increased 1.2 points.

Additional parallel refinement raised Gemini 3.5 Flash’s WorkBuddy-Bench score under OpenCode to 80.8.

Researchers also reported greater execution efficiency in three of four WorkBuddy-Bench configurations. Gemini under OpenCode reduced its average task steps from 25.4 to 21.9.

How Harness² Compares With Other AI Agent Optimization Methods

Harness² differs from approaches such as Meta-Harness, ReasoningBank, and Harness Scaling.

Meta-Harness searches for improved harness code using offline optimization, while ReasoningBank emphasizes retrieving stored experiences without modifying the underlying harness.

Harness Scaling modifies harness configurations for individual tasks but does not retain improvements across tasks.

Harness² combines task-specific adjustments with a persistent library of previous improvements.

In the researchers’ comparisons, Harness² achieved a WorkBuddy-Bench score of 77.5 using Gemini and OpenCode, compared with 76.4 for Meta-Harness, 73.4 for ReasoningBank, and 75.3 for Harness Scaling.

These comparative scores were measured in the Harness² study rather than independently reported by the competing projects.

What Harness² Means for AI Agent Development

Harness² demonstrates how improvements to prompts, tools, memory, and orchestration can increase AI agent performance without modifying the underlying model.

Its benchmark results suggest that optimizing the surrounding software infrastructure may offer meaningful gains, particularly for weaker agent configurations.

Although broader real-world validation remains important, Google’s open-source release provides developers with another approach to improving AI agent reliability, efficiency, and adaptability across professional workflows.