Today for AI
HOT RADAR
agentHEAT 0.0°

AI CURATED HOT STORY · 10/19/2026

Google Open-Sources Harness² for Recursive Agent Optimization

2 reports archived0 independent sourcesupdated 10/19/2026, 19:28:45
Synthesis & Latest Updates

Google Research has released and open-sourced Harness², a dual-layer recursive agent framework designed for open-world environments. This architecture addresses the critical challenges of overfitting and noise accumulation in scenarios lacking ground-truth answers by dynamically optimizing prompts, tools, and memory without altering model weights.

Harness² employs nested loops at both task and system levels: the former optimizes immediate execution components, while the latter accumulates improvement experiences to guide future tasks. Its Propose–Probe–Compose mechanism effectively distinguishes between general defect repair and domain-specific optimization. Empirical data shows that GLM-5, when equipped with this framework, saw an 11.5-point surge on WorkBuddy-Bench, significantly outperforming native Gemini 3.5 Flash, with performance gains reaching up to 11.8 percentage points after a single iteration.

This breakthrough marks a shift from static prompt engineering to dynamic self-evolution in agent development. Fully open-source and compatible with OpenCode and Codex, Harness² provides developers with a standardized scaffold to enhance model deployment capabilities at low cost, potentially accelerating scalable implementation across vertical AI applications.

LATEST/Harness² is now fully open-source and integrated into mainstream toolchains, with latest benchmarks showing it stably suppresses noise spirals in no-ground-truth environments and significantly boosts GLM-5's performance on complex office tasks after a single iteration.

Coverage timeline

Total 2 reports · Latest first
  1. Hacker News AIT2·85 pts

    Google Research Open-Sources Harness² for Recursive Agent Optimization

    • Introduces 'recursive agent harnessing' using a Propose-Probe-Compose step to optimize execution without ground-truth feedback
    • Employs a dual-level architecture: task-level refinement for current execution and system-level accumulation of lessons for future guidance
    • Delivers significant performance gains for GLM-5 and Gemini on benchmarks like WorkBuddy-Bench and JobBench with a single iteration
  2. 量子位官网T2·85 pts

    Google Open-Sources Harness² for Recursive Self-Improving Agents

    • Introduces a dual-layer recursive architecture (task-level + harnessing-level) to mitigate noise spirals in unsupervised agent scenarios.
    • Employs a Propose–Probe–Compose mechanism to distinguish between invariant defect fixes and domain-specific optimizations, enhancing scaffold generalization.
    • Fully open-sourced with support for OpenCode and Codex; benchmarks show GLM-5 scores increased by 11.5 points on WorkBuddy-Bench.