Harness2

Harness² is an open-source AI research framework developed by Google Research to improve autonomous agents through recursive optimization of their execution environments. Rather than modifying model weights, it refines the prompts, tools, memory, guardrails, and supporting components that determine agent behavior. Its task-level optimization follows a Propose–Probe–Compose process that tests alternative configurations and combines evidence from their performance. A second optimization level preserves effective configurations and lessons across successive tasks. The framework supports parallel exploration and sequential refinement without requiring ground-truth feedback. Researchers can experiment with OpenCode and Codex execution environments. It also integrates professional benchmarks for evaluating agent performance and studying improvements across different models and task categories.