Harness²: A Recursive Agent Harness Framework
Harness² is a plug-and-play framework that keeps model weights fixed, connects two levels of recursion, and exposes eight editable harness components.
Harness² is a plug-and-play open-source framework designed for recursive agent harnessing in open-world environments. Rather than fine-tuning model weights, the framework recursively refines the surrounding execution harness for open-world tasks.
Image source: https://github.com/google-research/harness2
Eight Modular Components
Harness² exposes eight distinct, editable components around the model:
- System prompt
- Guardrails
- Memory
- Tool guidance
- Subagents
- Skills
- Scripts
- Plugins
By treating these elements as editable layers, the framework connects task-level refinement with persistent improvement-harness updates.
Recursive Refinement and Learning
The framework operates by keeping model weights fixed while refining the execution harness. During execution, it proposes harness edits, probes their effects on the same task, and composes a refined harness from execution contrasts.
Adopted harnesses and evidence-supported lessons are retained in a persistent improvement harness to guide future refinement and tasks unseen during harness design.
Supported Executors and Benchmarks
Harness² provides built-in support for multiple execution modes, including parallel and sequential recursion, alongside integration with executors such as OpenCode and Codex. It supports JobBench, LAB, and WorkBuddy-Bench.
Model-backed benchmark runs require Node.js 20+, Google Cloud CLI, and benchmark-specific dependencies; WorkBuddy-Bench requires Docker.
Sources
- GitHub Repository: google-research/harness2
- X post: Alvaro Cintas on X