Google Cloud AI Research has open-sourced Harness², a new recursive agent harnessing framework designed to improve AI system performance without modifying frozen model weights. Released as version 0.1.0, the framework optimizes the code and files surrounding a model during deployment rather than retraining the underlying parameters. The system operates in the open world where tasks arrive unseen and without ground-truth verifiers, tackling agent optimization through runtime environmental adjustments.
The framework exposes eight editable components mounted as files, including system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. Harness² utilizes two levels of recursion to refine these elements. Task-level recursion employs a contrastive propose, probe, and compose step where an editor evaluates candidate improvements based on observed behavior. Harnessing-level recursion stores adopted changes and lessons in an experience file to guide future proposals.
In benchmark tests across WorkBuddy-Bench, JobBench, and Harvey's LAB using executors like Gemini and GLM-5, a single refinement step improved all tested model and harness setups. Notably, GLM-5 under OpenCode rose by 11.5 points on WorkBuddy-Bench, demonstrating that weaker executors can achieve significant gains using the framework. This recursive approach provides developers with a structured method to enhance agent capabilities at test time.