Google open-sources Harness², boosting GLM-5 benchmark scores by 11.5 points

Google Research has open-sourced Harness², an AI toolchain self-improvement framework, under the Apache 2.0 license—supporting OpenCode and Codex. The framework enables AI Agents to automatically refine system prompts, memory, skills, tool rules, scripts, and plugins—without retraining the underlying model. Its core mechanism relies on agent-driven trial-and-error: when confronted with a new task, the Agent first proposes multiple improvement strategies; then, different versions of the Agent execute each strategy independently and compare performance outcomes, ultimately synthesizing a more effective solution. Even in the absence of ground-truth answers, the system identifies improvement directions by analyzing real-world execution behavior. After task completion, it retains knowledge of which modifications worked—and why—enabling faster, more informed adaptation to similar future tasks. Google evaluated Harness² using GLM-5 and Gemini.